[Feat] add I2VGenXL for image-to-video generation (#6665)

---------
Co-authored-by: Dhruv Nair <dhruv.nair@gmail.com>
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
Co-authored-by: YiYi Xu <yixu310@gmail.com>

docs/source/en/_toctree.yml

@@ -284,6 +284,8 @@
       title: DiffEdit
     - local: api/pipelines/dit
       title: DiT
+    - local: api/pipelines/i2vgenxl
+      title: I2VGen-XL
     - local: api/pipelines/pix2pix
       title: InstructPix2Pix
     - local: api/pipelines/kandinsky

docs/source/en/api/pipelines/i2vgenxl.md

+<!--Copyright 2023 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+-->
+
+# I2VGen-XL
+
+[I2VGen-XL: High-Quality Image-to-Video Synthesis via Cascaded Diffusion Models](https://hf.co/papers/2311.04145.pdf) by Shiwei Zhang, Jiayu Wang, Yingya Zhang, Kang Zhao, Hangjie Yuan, Zhiwu Qin, Xiang Wang, Deli Zhao, and Jingren Zhou.
+
+The abstract from the paper is:
+
+*Video synthesis has recently made remarkable strides benefiting from the rapid development of diffusion models. However, it still encounters challenges in terms of semantic accuracy, clarity and spatio-temporal continuity. They primarily arise from the scarcity of well-aligned text-video data and the complex inherent structure of videos, making it difficult for the model to simultaneously ensure semantic and qualitative excellence. In this report, we propose a cascaded I2VGen-XL approach that enhances model performance by decoupling these two factors and ensures the alignment of the input data by utilizing static images as a form of crucial guidance. I2VGen-XL consists of two stages: i) the base stage guarantees coherent semantics and preserves content from input images by using two hierarchical encoders, and ii) the refinement stage enhances the video's details by incorporating an additional brief text and improves the resolution to 1280×720. To improve the diversity, we collect around 35 million single-shot text-video pairs and 6 billion text-image pairs to optimize the model. By this means, I2VGen-XL can simultaneously enhance the semantic accuracy, continuity of details and clarity of generated videos. Through extensive experiments, we have investigated the underlying principles of I2VGen-XL and compared it with current top methods, which can demonstrate its effectiveness on diverse data. The source code and models will be publicly available at [this https URL](https://i2vgen-xl.github.io/).*
+
+The original codebase can be found [here](https://github.com/ali-vilab/i2vgen-xl/). The model checkpoints can be found [here](https://huggingface.co/ali-vilab/). 
+
+<Tip>
+
+Make sure to check out the Schedulers [guide](../../using-diffusers/schedulers) to learn how to explore the tradeoff between scheduler speed and quality, and see the [reuse components across pipelines](../../using-diffusers/loading#reuse-components-across-pipelines) section to learn how to efficiently load the same components into multiple pipelines. Also, to know more about reducing the memory usage of this pipeline, refer to the ["Reduce memory usage"] section [here](../../using-diffusers/svd#reduce-memory-usage). 
+
+</Tip>
+
+Sample output with I2VGenXL:
+
+<table>
+    <tr>
+        <td><center>
+        masterpiece, bestquality, sunset.
+        <br>
+        <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/i2vgen-xl-example.gif"
+            alt="library"
+            style="width: 300px;" />
+        </center></td>
+    </tr>
+</table>
+
+## Notes
+
+* I2VGenXL always uses a `clip_skip` value of 1. This means it leverages the penultimate layer representations from the text encoder of CLIP.
+* It can generate videos of quality that is often on par with [Stable Video Diffusion](../../using-diffusers/svd) (SVD). 
+* Unlike SVD, it additionally accepts text prompts as inputs. 
+* It can generate higher resolution videos. 
+* When using the [`DDIMScheduler`] (which is default for this pipeline), less than 50 steps for inference leads to bad results.
+
+## I2VGenXLPipeline
+[[autodoc]] I2VGenXLPipeline
+	- all
+	- __call__
+
+## I2VGenXLPipelineOutput
+[[autodoc]] pipelines.i2vgen_xl.pipeline_i2vgen_xl.I2VGenXLPipelineOutput
\ No newline at end of file

scripts/convert_i2vgen_to_diffusers.py

+# coding=utf-8
+# Copyright 2023 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Conversion script for the LDM checkpoints. """
+
+import argparse
+
+import torch
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection
+
+from diffusers import DDIMScheduler, I2VGenXLPipeline, I2VGenXLUNet, StableDiffusionPipeline
+
+
+CLIP_ID = "laion/CLIP-ViT-H-14-laion2B-s32B-b79K"
+
+
+def assign_to_checkpoint(
+    paths, checkpoint, old_checkpoint, attention_paths_to_split=None, additional_replacements=None, config=None
+):
+    """
+    This does the final conversion step: take locally converted weights and apply a global renaming to them. It splits
+    attention layers, and takes into account additional replacements that may arise.
+
+    Assigns the weights to the new checkpoint.
+    """
+    assert isinstance(paths, list), "Paths should be a list of dicts containing 'old' and 'new' keys."
+
+    # Splits the attention layers into three variables.
+    if attention_paths_to_split is not None:
+        for path, path_map in attention_paths_to_split.items():
+            old_tensor = old_checkpoint[path]
+            channels = old_tensor.shape[0] // 3
+
+            target_shape = (-1, channels) if len(old_tensor.shape) == 3 else (-1)
+
+            num_heads = old_tensor.shape[0] // config["num_head_channels"] // 3
+
+            old_tensor = old_tensor.reshape((num_heads, 3 * channels // num_heads) + old_tensor.shape[1:])
+            query, key, value = old_tensor.split(channels // num_heads, dim=1)
+
+            checkpoint[path_map["query"]] = query.reshape(target_shape)
+            checkpoint[path_map["key"]] = key.reshape(target_shape)
+            checkpoint[path_map["value"]] = value.reshape(target_shape)
+
+    for path in paths:
+        new_path = path["new"]
+
+        # These have already been assigned
+        if attention_paths_to_split is not None and new_path in attention_paths_to_split:
+            continue
+
+        if additional_replacements is not None:
+            for replacement in additional_replacements:
+                new_path = new_path.replace(replacement["old"], replacement["new"])
+
+        # proj_attn.weight has to be converted from conv 1D to linear
+        weight = old_checkpoint[path["old"]]
+        names = ["proj_attn.weight"]
+        names_2 = ["proj_out.weight", "proj_in.weight"]
+        if any(k in new_path for k in names):
+            checkpoint[new_path] = weight[:, :, 0]
+        elif any(k in new_path for k in names_2) and len(weight.shape) > 2 and ".attentions." not in new_path:
+            checkpoint[new_path] = weight[:, :, 0]
+        else:
+            checkpoint[new_path] = weight
+
+
+def renew_attention_paths(old_list, n_shave_prefix_segments=0):
+    """
+    Updates paths inside attentions to the new naming scheme (local renaming)
+    """
+    mapping = []
+    for old_item in old_list:
+        new_item = old_item
+        mapping.append({"old": old_item, "new": new_item})
+
+    return mapping
+
+
+def shave_segments(path, n_shave_prefix_segments=1):
+    """
+    Removes segments. Positive values shave the first segments, negative shave the last segments.
+    """
+    if n_shave_prefix_segments >= 0:
+        return ".".join(path.split(".")[n_shave_prefix_segments:])
+    else:
+        return ".".join(path.split(".")[:n_shave_prefix_segments])
+
+
+def renew_temp_conv_paths(old_list, n_shave_prefix_segments=0):
+    """
+    Updates paths inside resnets to the new naming scheme (local renaming)
+    """
+    mapping = []
+    for old_item in old_list:
+        mapping.append({"old": old_item, "new": old_item})
+
+    return mapping
+
+
+def renew_resnet_paths(old_list, n_shave_prefix_segments=0):
+    """
+    Updates paths inside resnets to the new naming scheme (local renaming)
+    """
+    mapping = []
+    for old_item in old_list:
+        new_item = old_item.replace("in_layers.0", "norm1")
+        new_item = new_item.replace("in_layers.2", "conv1")
+
+        new_item = new_item.replace("out_layers.0", "norm2")
+        new_item = new_item.replace("out_layers.3", "conv2")
+
+        new_item = new_item.replace("emb_layers.1", "time_emb_proj")
+        new_item = new_item.replace("skip_connection", "conv_shortcut")
+
+        new_item = shave_segments(new_item, n_shave_prefix_segments=n_shave_prefix_segments)
+
+        if "temopral_conv" not in old_item:
+            mapping.append({"old": old_item, "new": new_item})
+
+    return mapping
+
+
+def convert_ldm_unet_checkpoint(checkpoint, config, path=None, extract_ema=False):
+    """
+    Takes a state dict and a config, and returns a converted checkpoint.
+    """
+
+    # extract state_dict for UNet
+    unet_state_dict = {}
+    keys = list(checkpoint.keys())
+
+    unet_key = "model.diffusion_model."
+
+    # at least a 100 parameters have to start with `model_ema` in order for the checkpoint to be EMA
+    if sum(k.startswith("model_ema") for k in keys) > 100 and extract_ema:
+        print(f"Checkpoint {path} has both EMA and non-EMA weights.")
+        print(
+            "In this conversion only the EMA weights are extracted. If you want to instead extract the non-EMA"
+            " weights (useful to continue fine-tuning), please make sure to remove the `--extract_ema` flag."
+        )
+        for key in keys:
+            if key.startswith("model.diffusion_model"):
+                flat_ema_key = "model_ema." + "".join(key.split(".")[1:])
+                unet_state_dict[key.replace(unet_key, "")] = checkpoint.pop(flat_ema_key)
+    else:
+        if sum(k.startswith("model_ema") for k in keys) > 100:
+            print(
+                "In this conversion only the non-EMA weights are extracted. If you want to instead extract the EMA"
+                " weights (usually better for inference), please make sure to add the `--extract_ema` flag."
+            )
+
+        for key in keys:
+            unet_state_dict[key.replace(unet_key, "")] = checkpoint.pop(key)
+
+    new_checkpoint = {}
+
+    new_checkpoint["time_embedding.linear_1.weight"] = unet_state_dict["time_embed.0.weight"]
+    new_checkpoint["time_embedding.linear_1.bias"] = unet_state_dict["time_embed.0.bias"]
+    new_checkpoint["time_embedding.linear_2.weight"] = unet_state_dict["time_embed.2.weight"]
+    new_checkpoint["time_embedding.linear_2.bias"] = unet_state_dict["time_embed.2.bias"]
+
+    additional_embedding_substrings = [
+        "local_image_concat",
+        "context_embedding",
+        "local_image_embedding",
+        "fps_embedding",
+    ]
+    for k in unet_state_dict:
+        if any(substring in k for substring in additional_embedding_substrings):
+            diffusers_key = k.replace("local_image_concat", "image_latents_proj_in").replace(
+                "local_image_embedding", "image_latents_context_embedding"
+            )
+            new_checkpoint[diffusers_key] = unet_state_dict[k]
+
+    # temporal encoder.
+    new_checkpoint["image_latents_temporal_encoder.norm1.weight"] = unet_state_dict[
+        "local_temporal_encoder.layers.0.0.norm.weight"
+    ]
+    new_checkpoint["image_latents_temporal_encoder.norm1.bias"] = unet_state_dict[
+        "local_temporal_encoder.layers.0.0.norm.bias"
+    ]
+
+    # attention
+    qkv = unet_state_dict["local_temporal_encoder.layers.0.0.fn.to_qkv.weight"]
+    q, k, v = torch.chunk(qkv, 3, dim=0)
+    new_checkpoint["image_latents_temporal_encoder.attn1.to_q.weight"] = q
+    new_checkpoint["image_latents_temporal_encoder.attn1.to_k.weight"] = k
+    new_checkpoint["image_latents_temporal_encoder.attn1.to_v.weight"] = v
+    new_checkpoint["image_latents_temporal_encoder.attn1.to_out.0.weight"] = unet_state_dict[
+        "local_temporal_encoder.layers.0.0.fn.to_out.0.weight"
+    ]
+    new_checkpoint["image_latents_temporal_encoder.attn1.to_out.0.bias"] = unet_state_dict[
+        "local_temporal_encoder.layers.0.0.fn.to_out.0.bias"
+    ]
+
+    # feedforward
+    new_checkpoint["image_latents_temporal_encoder.ff.net.0.proj.weight"] = unet_state_dict[
+        "local_temporal_encoder.layers.0.1.net.0.0.weight"
+    ]
+    new_checkpoint["image_latents_temporal_encoder.ff.net.0.proj.bias"] = unet_state_dict[
+        "local_temporal_encoder.layers.0.1.net.0.0.bias"
+    ]
+    new_checkpoint["image_latents_temporal_encoder.ff.net.2.weight"] = unet_state_dict[
+        "local_temporal_encoder.layers.0.1.net.2.weight"
+    ]
+    new_checkpoint["image_latents_temporal_encoder.ff.net.2.bias"] = unet_state_dict[
+        "local_temporal_encoder.layers.0.1.net.2.bias"
+    ]
+
+    if "class_embed_type" in config:
+        if config["class_embed_type"] is None:
+            # No parameters to port
+            ...
+        elif config["class_embed_type"] == "timestep" or config["class_embed_type"] == "projection":
+            new_checkpoint["class_embedding.linear_1.weight"] = unet_state_dict["label_emb.0.0.weight"]
+            new_checkpoint["class_embedding.linear_1.bias"] = unet_state_dict["label_emb.0.0.bias"]
+            new_checkpoint["class_embedding.linear_2.weight"] = unet_state_dict["label_emb.0.2.weight"]
+            new_checkpoint["class_embedding.linear_2.bias"] = unet_state_dict["label_emb.0.2.bias"]
+        else:
+            raise NotImplementedError(f"Not implemented `class_embed_type`: {config['class_embed_type']}")
+
+    new_checkpoint["conv_in.weight"] = unet_state_dict["input_blocks.0.0.weight"]
+    new_checkpoint["conv_in.bias"] = unet_state_dict["input_blocks.0.0.bias"]
+
+    first_temp_attention = [v for v in unet_state_dict if v.startswith("input_blocks.0.1")]
+    paths = renew_attention_paths(first_temp_attention)
+    meta_path = {"old": "input_blocks.0.1", "new": "transformer_in"}
+    assign_to_checkpoint(paths, new_checkpoint, unet_state_dict, additional_replacements=[meta_path], config=config)
+
+    new_checkpoint["conv_norm_out.weight"] = unet_state_dict["out.0.weight"]
+    new_checkpoint["conv_norm_out.bias"] = unet_state_dict["out.0.bias"]
+    new_checkpoint["conv_out.weight"] = unet_state_dict["out.2.weight"]
+    new_checkpoint["conv_out.bias"] = unet_state_dict["out.2.bias"]
+
+    # Retrieves the keys for the input blocks only
+    num_input_blocks = len({".".join(layer.split(".")[:2]) for layer in unet_state_dict if "input_blocks" in layer})
+    input_blocks = {
+        layer_id: [key for key in unet_state_dict if f"input_blocks.{layer_id}" in key]
+        for layer_id in range(num_input_blocks)
+    }
+
+    # Retrieves the keys for the middle blocks only
+    num_middle_blocks = len({".".join(layer.split(".")[:2]) for layer in unet_state_dict if "middle_block" in layer})
+    middle_blocks = {
+        layer_id: [key for key in unet_state_dict if f"middle_block.{layer_id}" in key]
+        for layer_id in range(num_middle_blocks)
+    }
+
+    # Retrieves the keys for the output blocks only
+    num_output_blocks = len({".".join(layer.split(".")[:2]) for layer in unet_state_dict if "output_blocks" in layer})
+    output_blocks = {
+        layer_id: [key for key in unet_state_dict if f"output_blocks.{layer_id}" in key]
+        for layer_id in range(num_output_blocks)
+    }
+
+    for i in range(1, num_input_blocks):
+        block_id = (i - 1) // (config["layers_per_block"] + 1)
+        layer_in_block_id = (i - 1) % (config["layers_per_block"] + 1)
+
+        resnets = [
+            key for key in input_blocks[i] if f"input_blocks.{i}.0" in key and f"input_blocks.{i}.0.op" not in key
+        ]
+        attentions = [key for key in input_blocks[i] if f"input_blocks.{i}.1" in key]
+        temp_attentions = [key for key in input_blocks[i] if f"input_blocks.{i}.2" in key]
+
+        if f"input_blocks.{i}.op.weight" in unet_state_dict:
+            new_checkpoint[f"down_blocks.{block_id}.downsamplers.0.conv.weight"] = unet_state_dict.pop(
+                f"input_blocks.{i}.op.weight"
+            )
+            new_checkpoint[f"down_blocks.{block_id}.downsamplers.0.conv.bias"] = unet_state_dict.pop(
+                f"input_blocks.{i}.op.bias"
+            )
+
+        paths = renew_resnet_paths(resnets)
+        meta_path = {"old": f"input_blocks.{i}.0", "new": f"down_blocks.{block_id}.resnets.{layer_in_block_id}"}
+        assign_to_checkpoint(
+            paths, new_checkpoint, unet_state_dict, additional_replacements=[meta_path], config=config
+        )
+
+        temporal_convs = [key for key in resnets if "temopral_conv" in key]
+        paths = renew_temp_conv_paths(temporal_convs)
+        meta_path = {
+            "old": f"input_blocks.{i}.0.temopral_conv",
+            "new": f"down_blocks.{block_id}.temp_convs.{layer_in_block_id}",
+        }
+        assign_to_checkpoint(
+            paths, new_checkpoint, unet_state_dict, additional_replacements=[meta_path], config=config
+        )
+
+        if len(attentions):
+            paths = renew_attention_paths(attentions)
+            meta_path = {"old": f"input_blocks.{i}.1", "new": f"down_blocks.{block_id}.attentions.{layer_in_block_id}"}
+            assign_to_checkpoint(
+                paths, new_checkpoint, unet_state_dict, additional_replacements=[meta_path], config=config
+            )
+
+        if len(temp_attentions):
+            paths = renew_attention_paths(temp_attentions)
+            meta_path = {
+                "old": f"input_blocks.{i}.2",
+                "new": f"down_blocks.{block_id}.temp_attentions.{layer_in_block_id}",
+            }
+            assign_to_checkpoint(
+                paths, new_checkpoint, unet_state_dict, additional_replacements=[meta_path], config=config
+            )
+
+    resnet_0 = middle_blocks[0]
+    temporal_convs_0 = [key for key in resnet_0 if "temopral_conv" in key]
+    attentions = middle_blocks[1]
+    temp_attentions = middle_blocks[2]
+    resnet_1 = middle_blocks[3]
+    temporal_convs_1 = [key for key in resnet_1 if "temopral_conv" in key]
+
+    resnet_0_paths = renew_resnet_paths(resnet_0)
+    meta_path = {"old": "middle_block.0", "new": "mid_block.resnets.0"}
+    assign_to_checkpoint(
+        resnet_0_paths, new_checkpoint, unet_state_dict, config=config, additional_replacements=[meta_path]
+    )
+
+    temp_conv_0_paths = renew_temp_conv_paths(temporal_convs_0)
+    meta_path = {"old": "middle_block.0.temopral_conv", "new": "mid_block.temp_convs.0"}
+    assign_to_checkpoint(
+        temp_conv_0_paths, new_checkpoint, unet_state_dict, config=config, additional_replacements=[meta_path]
+    )
+
+    resnet_1_paths = renew_resnet_paths(resnet_1)
+    meta_path = {"old": "middle_block.3", "new": "mid_block.resnets.1"}
+    assign_to_checkpoint(
+        resnet_1_paths, new_checkpoint, unet_state_dict, config=config, additional_replacements=[meta_path]
+    )
+
+    temp_conv_1_paths = renew_temp_conv_paths(temporal_convs_1)
+    meta_path = {"old": "middle_block.3.temopral_conv", "new": "mid_block.temp_convs.1"}
+    assign_to_checkpoint(
+        temp_conv_1_paths, new_checkpoint, unet_state_dict, config=config, additional_replacements=[meta_path]
+    )
+
+    attentions_paths = renew_attention_paths(attentions)
+    meta_path = {"old": "middle_block.1", "new": "mid_block.attentions.0"}
+    assign_to_checkpoint(
+        attentions_paths, new_checkpoint, unet_state_dict, additional_replacements=[meta_path], config=config
+    )
+
+    temp_attentions_paths = renew_attention_paths(temp_attentions)
+    meta_path = {"old": "middle_block.2", "new": "mid_block.temp_attentions.0"}
+    assign_to_checkpoint(
+        temp_attentions_paths, new_checkpoint, unet_state_dict, additional_replacements=[meta_path], config=config
+    )
+
+    for i in range(num_output_blocks):
+        block_id = i // (config["layers_per_block"] + 1)
+        layer_in_block_id = i % (config["layers_per_block"] + 1)
+        output_block_layers = [shave_segments(name, 2) for name in output_blocks[i]]
+        output_block_list = {}
+
+        for layer in output_block_layers:
+            layer_id, layer_name = layer.split(".")[0], shave_segments(layer, 1)
+            if layer_id in output_block_list:
+                output_block_list[layer_id].append(layer_name)
+            else:
+                output_block_list[layer_id] = [layer_name]
+
+        if len(output_block_list) > 1:
+            resnets = [key for key in output_blocks[i] if f"output_blocks.{i}.0" in key]
+            attentions = [key for key in output_blocks[i] if f"output_blocks.{i}.1" in key]
+            temp_attentions = [key for key in output_blocks[i] if f"output_blocks.{i}.2" in key]
+
+            resnet_0_paths = renew_resnet_paths(resnets)
+            paths = renew_resnet_paths(resnets)
+
+            meta_path = {"old": f"output_blocks.{i}.0", "new": f"up_blocks.{block_id}.resnets.{layer_in_block_id}"}
+            assign_to_checkpoint(
+                paths, new_checkpoint, unet_state_dict, additional_replacements=[meta_path], config=config
+            )
+
+            temporal_convs = [key for key in resnets if "temopral_conv" in key]
+            paths = renew_temp_conv_paths(temporal_convs)
+            meta_path = {
+                "old": f"output_blocks.{i}.0.temopral_conv",
+                "new": f"up_blocks.{block_id}.temp_convs.{layer_in_block_id}",
+            }
+            assign_to_checkpoint(
+                paths, new_checkpoint, unet_state_dict, additional_replacements=[meta_path], config=config
+            )
+
+            output_block_list = {k: sorted(v) for k, v in output_block_list.items()}
+            if ["conv.bias", "conv.weight"] in output_block_list.values():
+                index = list(output_block_list.values()).index(["conv.bias", "conv.weight"])
+                new_checkpoint[f"up_blocks.{block_id}.upsamplers.0.conv.weight"] = unet_state_dict[
+                    f"output_blocks.{i}.{index}.conv.weight"
+                ]
+                new_checkpoint[f"up_blocks.{block_id}.upsamplers.0.conv.bias"] = unet_state_dict[
+                    f"output_blocks.{i}.{index}.conv.bias"
+                ]
+
+                # Clear attentions as they have been attributed above.
+                if len(attentions) == 2:
+                    attentions = []
+
+            if len(attentions):
+                paths = renew_attention_paths(attentions)
+                meta_path = {
+                    "old": f"output_blocks.{i}.1",
+                    "new": f"up_blocks.{block_id}.attentions.{layer_in_block_id}",
+                }
+                assign_to_checkpoint(
+                    paths, new_checkpoint, unet_state_dict, additional_replacements=[meta_path], config=config
+                )
+
+            if len(temp_attentions):
+                paths = renew_attention_paths(temp_attentions)
+                meta_path = {
+                    "old": f"output_blocks.{i}.2",
+                    "new": f"up_blocks.{block_id}.temp_attentions.{layer_in_block_id}",
+                }
+                assign_to_checkpoint(
+                    paths, new_checkpoint, unet_state_dict, additional_replacements=[meta_path], config=config
+                )
+        else:
+            resnet_0_paths = renew_resnet_paths(output_block_layers, n_shave_prefix_segments=1)
+            for path in resnet_0_paths:
+                old_path = ".".join(["output_blocks", str(i), path["old"]])
+                new_path = ".".join(["up_blocks", str(block_id), "resnets", str(layer_in_block_id), path["new"]])
+                new_checkpoint[new_path] = unet_state_dict[old_path]
+
+            temopral_conv_paths = [l for l in output_block_layers if "temopral_conv" in l]
+            for path in temopral_conv_paths:
+                pruned_path = path.split("temopral_conv.")[-1]
+                old_path = ".".join(["output_blocks", str(i), str(block_id), "temopral_conv", pruned_path])
+                new_path = ".".join(["up_blocks", str(block_id), "temp_convs", str(layer_in_block_id), pruned_path])
+                new_checkpoint[new_path] = unet_state_dict[old_path]
+
+    return new_checkpoint
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--unet_checkpoint_path", default=None, type=str, required=True, help="Path to the checkpoint to convert."
+    )
+    parser.add_argument("--dump_path", default=None, type=str, required=True, help="Path to the output model.")
+    parser.add_argument("--push_to_hub", action="store_true")
+    args = parser.parse_args()
+
+    # UNet
+    unet_checkpoint = torch.load(args.unet_checkpoint_path, map_location="cpu")
+    unet_checkpoint = unet_checkpoint["state_dict"]
+    unet = I2VGenXLUNet(sample_size=32)
+
+    converted_ckpt = convert_ldm_unet_checkpoint(unet_checkpoint, unet.config)
+
+    diff_0 = set(unet.state_dict().keys()) - set(converted_ckpt.keys())
+    diff_1 = set(converted_ckpt.keys()) - set(unet.state_dict().keys())
+
+    assert len(diff_0) == len(diff_1) == 0, "Converted weights don't match"
+
+    unet.load_state_dict(converted_ckpt, strict=True)
+
+    # vae
+    temp_pipe = StableDiffusionPipeline.from_single_file(
+        "https://huggingface.co/ali-vilab/i2vgen-xl/blob/main/models/v2-1_512-ema-pruned.ckpt"
+    )
+    vae = temp_pipe.vae
+    del temp_pipe
+
+    # text encoder and tokenizer
+    text_encoder = CLIPTextModel.from_pretrained(CLIP_ID)
+    tokenizer = CLIPTokenizer.from_pretrained(CLIP_ID)
+
+    # image encoder and feature extractor
+    image_encoder = CLIPVisionModelWithProjection.from_pretrained(CLIP_ID)
+    feature_extractor = CLIPImageProcessor.from_pretrained(CLIP_ID)
+
+    # scheduler
+    # https://github.com/ali-vilab/i2vgen-xl/blob/main/configs/i2vgen_xl_train.yaml
+    scheduler = DDIMScheduler(
+        beta_schedule="squaredcos_cap_v2",
+        rescale_betas_zero_snr=True,
+        set_alpha_to_one=True,
+        clip_sample=False,
+        steps_offset=1,
+        timestep_spacing="leading",
+        prediction_type="v_prediction",
+    )
+
+    # final
+    pipeline = I2VGenXLPipeline(
+        unet=unet,
+        vae=vae,
+        image_encoder=image_encoder,
+        feature_extractor=feature_extractor,
+        text_encoder=text_encoder,
+        tokenizer=tokenizer,
+        scheduler=scheduler,
+    )
+
+    pipeline.save_pretrained(args.dump_path, push_to_hub=args.push_to_hub)

src/diffusers/__init__.py

@@ -80,6 +80,7 @@ else:
             "AutoencoderTiny",
             "ConsistencyDecoderVAE",
             "ControlNetModel",
+            "I2VGenXLUNet",
             "Kandinsky3UNet",
             "ModelMixin",
             "MotionAdapter",
@@ -217,6 +218,7 @@ else:
             "BlipDiffusionPipeline",
             "CLIPImageProjection",
             "CycleDiffusionPipeline",
+            "I2VGenXLPipeline",
             "IFImg2ImgPipeline",
             "IFImg2ImgSuperResolutionPipeline",
             "IFInpaintingPipeline",
@@ -462,6 +464,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             AutoencoderTiny,
             ConsistencyDecoderVAE,
             ControlNetModel,
+            I2VGenXLUNet,
             Kandinsky3UNet,
             ModelMixin,
             MotionAdapter,
@@ -578,6 +581,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             AudioLDMPipeline,
             CLIPImageProjection,
             CycleDiffusionPipeline,
+            I2VGenXLPipeline,
             IFImg2ImgPipeline,
             IFImg2ImgSuperResolutionPipeline,
             IFInpaintingPipeline,

src/diffusers/models/__init__.py

@@ -43,6 +43,7 @@ if is_torch_available():
     _import_structure["unets.unet_2d"] = ["UNet2DModel"]
     _import_structure["unets.unet_2d_condition"] = ["UNet2DConditionModel"]
     _import_structure["unets.unet_3d_condition"] = ["UNet3DConditionModel"]
+    _import_structure["unets.unet_i2vgen_xl"] = ["I2VGenXLUNet"]
     _import_structure["unets.unet_kandinsky3"] = ["Kandinsky3UNet"]
     _import_structure["unets.unet_motion_model"] = ["MotionAdapter", "UNetMotionModel"]
     _import_structure["unets.unet_spatio_temporal_condition"] = ["UNetSpatioTemporalConditionModel"]
@@ -76,6 +77,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             TransformerTemporalModel,
         )
         from .unets import (
+            I2VGenXLUNet,
             Kandinsky3UNet,
             MotionAdapter,
             UNet1DModel,

src/diffusers/models/attention.py

@@ -143,7 +143,7 @@ class BasicTransformerBlock(nn.Module):
         double_self_attention: bool = False,
         upcast_attention: bool = False,
         norm_elementwise_affine: bool = True,
-        norm_type: str = "layer_norm",  # 'layer_norm', 'ada_norm', 'ada_norm_zero', 'ada_norm_single'
+        norm_type: str = "layer_norm",  # 'layer_norm', 'ada_norm', 'ada_norm_zero', 'ada_norm_single', 'layer_norm_i2vgen'
         norm_eps: float = 1e-5,
         final_dropout: bool = False,
         attention_type: str = "default",
@@ -158,18 +158,15 @@ class BasicTransformerBlock(nn.Module):
         super().__init__()
         self.only_cross_attention = only_cross_attention
 
-        self.use_ada_layer_norm_zero = (num_embeds_ada_norm is not None) and norm_type == "ada_norm_zero"
-        self.use_ada_layer_norm = (num_embeds_ada_norm is not None) and norm_type == "ada_norm"
-        self.use_ada_layer_norm_single = norm_type == "ada_norm_single"
-        self.use_layer_norm = norm_type == "layer_norm"
-        self.use_ada_layer_norm_continuous = norm_type == "ada_norm_continuous"
-
         if norm_type in ("ada_norm", "ada_norm_zero") and num_embeds_ada_norm is None:
             raise ValueError(
                 f"`norm_type` is set to {norm_type}, but `num_embeds_ada_norm` is not defined. Please make sure to"
                 f" define `num_embeds_ada_norm` if setting `norm_type` to {norm_type}."
             )
 
+        self.norm_type = norm_type
+        self.num_embeds_ada_norm = num_embeds_ada_norm
+
         if positional_embeddings and (num_positional_embeddings is None):
             raise ValueError(
                 "If `positional_embedding` type is defined, `num_positition_embeddings` must also be defined."
@@ -182,11 +179,11 @@ class BasicTransformerBlock(nn.Module):
 
         # Define 3 blocks. Each block has its own normalization layer.
         # 1. Self-Attn
-        if self.use_ada_layer_norm:
+        if norm_type == "ada_norm":
             self.norm1 = AdaLayerNorm(dim, num_embeds_ada_norm)
-        elif self.use_ada_layer_norm_zero:
+        elif norm_type == "ada_norm_zero":
             self.norm1 = AdaLayerNormZero(dim, num_embeds_ada_norm)
-        elif self.use_ada_layer_norm_continuous:
+        elif norm_type == "ada_norm_continuous":
             self.norm1 = AdaLayerNormContinuous(
                 dim,
                 ada_norm_continous_conditioning_embedding_dim,
@@ -214,9 +211,9 @@ class BasicTransformerBlock(nn.Module):
             # We currently only use AdaLayerNormZero for self attention where there will only be one attention block.
             # I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during
             # the second cross attention block.
-            if self.use_ada_layer_norm:
+            if norm_type == "ada_norm":
                 self.norm2 = AdaLayerNorm(dim, num_embeds_ada_norm)
-            elif self.use_ada_layer_norm_continuous:
+            elif norm_type == "ada_norm_continuous":
                 self.norm2 = AdaLayerNormContinuous(
                     dim,
                     ada_norm_continous_conditioning_embedding_dim,
@@ -243,7 +240,7 @@ class BasicTransformerBlock(nn.Module):
             self.attn2 = None
 
         # 3. Feed-forward
-        if self.use_ada_layer_norm_continuous:
+        if norm_type == "ada_norm_continuous":
             self.norm3 = AdaLayerNormContinuous(
                 dim,
                 ada_norm_continous_conditioning_embedding_dim,
@@ -252,8 +249,11 @@ class BasicTransformerBlock(nn.Module):
                 ada_norm_bias,
                 "layer_norm",
             )
-        elif not self.use_ada_layer_norm_single:
+
+        elif norm_type in ["ada_norm_zero", "ada_norm", "layer_norm", "ada_norm_continuous"]:
             self.norm3 = nn.LayerNorm(dim, norm_eps, norm_elementwise_affine)
+        elif norm_type == "layer_norm_i2vgen":
+            self.norm3 = None
 
         self.ff = FeedForward(
             dim,
@@ -269,7 +269,7 @@ class BasicTransformerBlock(nn.Module):
             self.fuser = GatedSelfAttentionDense(dim, cross_attention_dim, num_attention_heads, attention_head_dim)
 
         # 5. Scale-shift for PixArt-Alpha.
-        if self.use_ada_layer_norm_single:
+        if norm_type == "ada_norm_single":
             self.scale_shift_table = nn.Parameter(torch.randn(6, dim) / dim**0.5)
 
         # let chunk size default to None
@@ -296,17 +296,17 @@ class BasicTransformerBlock(nn.Module):
         # 0. Self-Attention
         batch_size = hidden_states.shape[0]
 
-        if self.use_ada_layer_norm:
+        if self.norm_type == "ada_norm":
             norm_hidden_states = self.norm1(hidden_states, timestep)
-        elif self.use_ada_layer_norm_zero:
+        elif self.norm_type == "ada_norm_zero":
             norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1(
                 hidden_states, timestep, class_labels, hidden_dtype=hidden_states.dtype
             )
-        elif self.use_layer_norm:
+        elif self.norm_type in ["layer_norm", "layer_norm_i2vgen"]:
             norm_hidden_states = self.norm1(hidden_states)
-        elif self.use_ada_layer_norm_continuous:
+        elif self.norm_type == "ada_norm_continuous":
             norm_hidden_states = self.norm1(hidden_states, added_cond_kwargs["pooled_text_emb"])
-        elif self.use_ada_layer_norm_single:
+        elif self.norm_type == "ada_norm_single":
             shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (
                 self.scale_shift_table[None] + timestep.reshape(batch_size, 6, -1)
             ).chunk(6, dim=1)
@@ -332,9 +332,9 @@ class BasicTransformerBlock(nn.Module):
             attention_mask=attention_mask,
             **cross_attention_kwargs,
         )
-        if self.use_ada_layer_norm_zero:
+        if self.norm_type == "ada_norm_zero":
             attn_output = gate_msa.unsqueeze(1) * attn_output
-        elif self.use_ada_layer_norm_single:
+        elif self.norm_type == "ada_norm_single":
             attn_output = gate_msa * attn_output
 
         hidden_states = attn_output + hidden_states
@@ -347,20 +347,20 @@ class BasicTransformerBlock(nn.Module):
 
         # 3. Cross-Attention
         if self.attn2 is not None:
-            if self.use_ada_layer_norm:
+            if self.norm_type == "ada_norm":
                 norm_hidden_states = self.norm2(hidden_states, timestep)
-            elif self.use_ada_layer_norm_zero or self.use_layer_norm:
+            elif self.norm_type in ["ada_norm_zero", "layer_norm", "layer_norm_i2vgen"]:
                 norm_hidden_states = self.norm2(hidden_states)
-            elif self.use_ada_layer_norm_single:
+            elif self.norm_type == "ada_norm_single":
                 # For PixArt norm2 isn't applied here:
                 # https://github.com/PixArt-alpha/PixArt-alpha/blob/0f55e922376d8b797edd44d25d0e7464b260dcab/diffusion/model/nets/PixArtMS.py#L70C1-L76C103
                 norm_hidden_states = hidden_states
-            elif self.use_ada_layer_norm_continuous:
+            elif self.norm_type == "ada_norm_continuous":
                 norm_hidden_states = self.norm2(hidden_states, added_cond_kwargs["pooled_text_emb"])
             else:
                 raise ValueError("Incorrect norm")
 
-            if self.pos_embed is not None and self.use_ada_layer_norm_single is False:
+            if self.pos_embed is not None and self.norm_type != "ada_norm_single":
                 norm_hidden_states = self.pos_embed(norm_hidden_states)
 
             attn_output = self.attn2(
@@ -372,15 +372,16 @@ class BasicTransformerBlock(nn.Module):
             hidden_states = attn_output + hidden_states
 
         # 4. Feed-forward
-        if self.use_ada_layer_norm_continuous:
+        # i2vgen doesn't have this norm 🤷‍♂️
+        if self.norm_type == "ada_norm_continuous":
             norm_hidden_states = self.norm3(hidden_states, added_cond_kwargs["pooled_text_emb"])
-        elif not self.use_ada_layer_norm_single:
+        elif not self.norm_type == "ada_norm_single":
             norm_hidden_states = self.norm3(hidden_states)
 
-        if self.use_ada_layer_norm_zero:
+        if self.norm_type == "ada_norm_zero":
             norm_hidden_states = norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None]
 
-        if self.use_ada_layer_norm_single:
+        if self.norm_type == "ada_norm_single":
             norm_hidden_states = self.norm2(hidden_states)
             norm_hidden_states = norm_hidden_states * (1 + scale_mlp) + shift_mlp
 
@@ -392,9 +393,9 @@ class BasicTransformerBlock(nn.Module):
         else:
             ff_output = self.ff(norm_hidden_states, scale=lora_scale)
 
-        if self.use_ada_layer_norm_zero:
+        if self.norm_type == "ada_norm_zero":
             ff_output = gate_mlp.unsqueeze(1) * ff_output
-        elif self.use_ada_layer_norm_single:
+        elif self.norm_type == "ada_norm_single":
             ff_output = gate_mlp * ff_output
 
         hidden_states = ff_output + hidden_states

src/diffusers/models/unets/__init__.py

@@ -6,6 +6,7 @@ if is_torch_available():
     from .unet_2d import UNet2DModel
     from .unet_2d_condition import UNet2DConditionModel
     from .unet_3d_condition import UNet3DConditionModel
+    from .unet_i2vgen_xl import I2VGenXLUNet
     from .unet_kandinsky3 import Kandinsky3UNet
     from .unet_motion_model import MotionAdapter, UNetMotionModel
     from .unet_spatio_temporal_condition import UNetSpatioTemporalConditionModel

src/diffusers/models/unets/unet_i2vgen_xl.py

+# Copyright 2023 Alibaba DAMO-VILAB and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Any, Dict, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.utils.checkpoint
+
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...loaders import UNet2DConditionLoadersMixin
+from ...utils import logging
+from ..activations import get_activation
+from ..attention import Attention, FeedForward
+from ..attention_processor import (
+    ADDED_KV_ATTENTION_PROCESSORS,
+    CROSS_ATTENTION_PROCESSORS,
+    AttentionProcessor,
+    AttnAddedKVProcessor,
+    AttnProcessor,
+)
+from ..embeddings import TimestepEmbedding, Timesteps
+from ..modeling_utils import ModelMixin
+from ..transformers.transformer_temporal import TransformerTemporalModel
+from .unet_3d_blocks import (
+    CrossAttnDownBlock3D,
+    CrossAttnUpBlock3D,
+    DownBlock3D,
+    UNetMidBlock3DCrossAttn,
+    UpBlock3D,
+    get_down_block,
+    get_up_block,
+)
+from .unet_3d_condition import UNet3DConditionOutput
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+def _to_tensor(inputs, device):
+    if not torch.is_tensor(inputs):
+        # TODO: this requires sync between CPU and GPU. So try to pass `inputs` as tensors if you can
+        # This would be a good case for the `match` statement (Python 3.10+)
+        is_mps = device.type == "mps"
+        if isinstance(inputs, float):
+            dtype = torch.float32 if is_mps else torch.float64
+        else:
+            dtype = torch.int32 if is_mps else torch.int64
+        inputs = torch.tensor([inputs], dtype=dtype, device=device)
+    elif len(inputs.shape) == 0:
+        inputs = inputs[None].to(device)
+
+    return inputs
+
+
+def _collapse_frames_into_batch(sample: torch.Tensor) -> torch.Tensor:
+    batch_size, channels, num_frames, height, width = sample.shape
+    sample = sample.permute(0, 2, 1, 3, 4).reshape(batch_size * num_frames, channels, height, width)
+
+    return sample
+
+
+class I2VGenXLTransformerTemporalEncoder(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        activation_fn: str = "geglu",
+        upcast_attention: bool = False,
+        ff_inner_dim: Optional[int] = None,
+        dropout: int = 0.0,
+    ):
+        super().__init__()
+        self.norm1 = nn.LayerNorm(dim, elementwise_affine=True, eps=1e-5)
+        self.attn1 = Attention(
+            query_dim=dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            dropout=dropout,
+            bias=False,
+            upcast_attention=upcast_attention,
+            out_bias=True,
+        )
+        self.ff = FeedForward(
+            dim,
+            dropout=dropout,
+            activation_fn=activation_fn,
+            final_dropout=False,
+            inner_dim=ff_inner_dim,
+            bias=True,
+        )
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+    ) -> torch.FloatTensor:
+        norm_hidden_states = self.norm1(hidden_states)
+        attn_output = self.attn1(norm_hidden_states, encoder_hidden_states=None)
+        hidden_states = attn_output + hidden_states
+        if hidden_states.ndim == 4:
+            hidden_states = hidden_states.squeeze(1)
+
+        ff_output = self.ff(hidden_states, scale=1.0)
+        hidden_states = ff_output + hidden_states
+        if hidden_states.ndim == 4:
+            hidden_states = hidden_states.squeeze(1)
+
+        return hidden_states
+
+
+class I2VGenXLUNet(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin):
+    r"""
+    I2VGenXL UNet. It is a conditional 3D UNet model that takes a noisy sample, conditional state, and a timestep
+    and returns a sample-shaped output.
+
+    This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented
+    for all models (such as downloading or saving).
+
+    Parameters:
+        sample_size (`int` or `Tuple[int, int]`, *optional*, defaults to `None`):
+            Height and width of input/output sample.
+        in_channels (`int`, *optional*, defaults to 4): The number of channels in the input sample.
+        out_channels (`int`, *optional*, defaults to 4): The number of channels in the output.
+        down_block_types (`Tuple[str]`, *optional*, defaults to `("CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "DownBlock2D")`):
+            The tuple of downsample blocks to use.
+        up_block_types (`Tuple[str]`, *optional*, defaults to `("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D")`):
+            The tuple of upsample blocks to use.
+        block_out_channels (`Tuple[int]`, *optional*, defaults to `(320, 640, 1280, 1280)`):
+            The tuple of output channels for each block.
+        layers_per_block (`int`, *optional*, defaults to 2): The number of layers per block.
+        norm_num_groups (`int`, *optional*, defaults to 32): The number of groups to use for the normalization.
+            If `None`, normalization and activation layers is skipped in post-processing.
+        cross_attention_dim (`int`, *optional*, defaults to 1280): The dimension of the cross attention features.
+        num_attention_heads (`int`, *optional*): The number of attention heads.
+    """
+
+    _supports_gradient_checkpointing = False
+
+    @register_to_config
+    def __init__(
+        self,
+        sample_size: Optional[int] = None,
+        in_channels: int = 4,
+        out_channels: int = 4,
+        down_block_types: Tuple[str, ...] = (
+            "CrossAttnDownBlock3D",
+            "CrossAttnDownBlock3D",
+            "CrossAttnDownBlock3D",
+            "DownBlock3D",
+        ),
+        up_block_types: Tuple[str, ...] = (
+            "UpBlock3D",
+            "CrossAttnUpBlock3D",
+            "CrossAttnUpBlock3D",
+            "CrossAttnUpBlock3D",
+        ),
+        block_out_channels: Tuple[int, ...] = (320, 640, 1280, 1280),
+        layers_per_block: int = 2,
+        norm_num_groups: Optional[int] = 32,
+        cross_attention_dim: int = 1024,
+        num_attention_heads: Optional[Union[int, Tuple[int]]] = 64,
+    ):
+        super().__init__()
+
+        self.sample_size = sample_size
+
+        # Check inputs
+        if len(down_block_types) != len(up_block_types):
+            raise ValueError(
+                f"Must provide the same number of `down_block_types` as `up_block_types`. `down_block_types`: {down_block_types}. `up_block_types`: {up_block_types}."
+            )
+
+        if len(block_out_channels) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `block_out_channels` as `down_block_types`. `block_out_channels`: {block_out_channels}. `down_block_types`: {down_block_types}."
+            )
+
+        if not isinstance(num_attention_heads, int) and len(num_attention_heads) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `num_attention_heads` as `down_block_types`. `num_attention_heads`: {num_attention_heads}. `down_block_types`: {down_block_types}."
+            )
+
+        # input
+        self.conv_in = nn.Conv2d(in_channels + in_channels, block_out_channels[0], kernel_size=3, padding=1)
+
+        self.transformer_in = TransformerTemporalModel(
+            num_attention_heads=8,
+            attention_head_dim=num_attention_heads,
+            in_channels=block_out_channels[0],
+            num_layers=1,
+            norm_num_groups=norm_num_groups,
+        )
+
+        # image embedding
+        self.image_latents_proj_in = nn.Sequential(
+            nn.Conv2d(4, in_channels * 4, 3, padding=1),
+            nn.SiLU(),
+            nn.Conv2d(in_channels * 4, in_channels * 4, 3, stride=1, padding=1),
+            nn.SiLU(),
+            nn.Conv2d(in_channels * 4, in_channels, 3, stride=1, padding=1),
+        )
+        self.image_latents_temporal_encoder = I2VGenXLTransformerTemporalEncoder(
+            dim=in_channels,
+            num_attention_heads=2,
+            ff_inner_dim=in_channels * 4,
+            attention_head_dim=in_channels,
+            activation_fn="gelu",
+        )
+        self.image_latents_context_embedding = nn.Sequential(
+            nn.Conv2d(4, in_channels * 8, 3, padding=1),
+            nn.SiLU(),
+            nn.AdaptiveAvgPool2d((32, 32)),
+            nn.Conv2d(in_channels * 8, in_channels * 16, 3, stride=2, padding=1),
+            nn.SiLU(),
+            nn.Conv2d(in_channels * 16, cross_attention_dim, 3, stride=2, padding=1),
+        )
+
+        # other embeddings -- time, context, fps, etc.
+        time_embed_dim = block_out_channels[0] * 4
+        self.time_proj = Timesteps(block_out_channels[0], True, 0)
+        timestep_input_dim = block_out_channels[0]
+
+        self.time_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim, act_fn="silu")
+        self.context_embedding = nn.Sequential(
+            nn.Linear(cross_attention_dim, time_embed_dim),
+            nn.SiLU(),
+            nn.Linear(time_embed_dim, cross_attention_dim * in_channels),
+        )
+        self.fps_embedding = nn.Sequential(
+            nn.Linear(timestep_input_dim, time_embed_dim), nn.SiLU(), nn.Linear(time_embed_dim, time_embed_dim)
+        )
+
+        # blocks
+        self.down_blocks = nn.ModuleList([])
+        self.up_blocks = nn.ModuleList([])
+
+        if isinstance(num_attention_heads, int):
+            num_attention_heads = (num_attention_heads,) * len(down_block_types)
+
+        # down
+        output_channel = block_out_channels[0]
+        for i, down_block_type in enumerate(down_block_types):
+            input_channel = output_channel
+            output_channel = block_out_channels[i]
+            is_final_block = i == len(block_out_channels) - 1
+
+            down_block = get_down_block(
+                down_block_type,
+                num_layers=layers_per_block,
+                in_channels=input_channel,
+                out_channels=output_channel,
+                temb_channels=time_embed_dim,
+                add_downsample=not is_final_block,
+                resnet_eps=1e-05,
+                resnet_act_fn="silu",
+                resnet_groups=norm_num_groups,
+                cross_attention_dim=cross_attention_dim,
+                num_attention_heads=num_attention_heads[i],
+                downsample_padding=1,
+                dual_cross_attention=False,
+            )
+            self.down_blocks.append(down_block)
+
+        # mid
+        self.mid_block = UNetMidBlock3DCrossAttn(
+            in_channels=block_out_channels[-1],
+            temb_channels=time_embed_dim,
+            resnet_eps=1e-05,
+            resnet_act_fn="silu",
+            output_scale_factor=1,
+            cross_attention_dim=cross_attention_dim,
+            num_attention_heads=num_attention_heads[-1],
+            resnet_groups=norm_num_groups,
+            dual_cross_attention=False,
+        )
+
+        # count how many layers upsample the images
+        self.num_upsamplers = 0
+
+        # up
+        reversed_block_out_channels = list(reversed(block_out_channels))
+        reversed_num_attention_heads = list(reversed(num_attention_heads))
+
+        output_channel = reversed_block_out_channels[0]
+        for i, up_block_type in enumerate(up_block_types):
+            is_final_block = i == len(block_out_channels) - 1
+
+            prev_output_channel = output_channel
+            output_channel = reversed_block_out_channels[i]
+            input_channel = reversed_block_out_channels[min(i + 1, len(block_out_channels) - 1)]
+
+            # add upsample block for all BUT final layer
+            if not is_final_block:
+                add_upsample = True
+                self.num_upsamplers += 1
+            else:
+                add_upsample = False
+
+            up_block = get_up_block(
+                up_block_type,
+                num_layers=layers_per_block + 1,
+                in_channels=input_channel,
+                out_channels=output_channel,
+                prev_output_channel=prev_output_channel,
+                temb_channels=time_embed_dim,
+                add_upsample=add_upsample,
+                resnet_eps=1e-05,
+                resnet_act_fn="silu",
+                resnet_groups=norm_num_groups,
+                cross_attention_dim=cross_attention_dim,
+                num_attention_heads=reversed_num_attention_heads[i],
+                dual_cross_attention=False,
+                resolution_idx=i,
+            )
+            self.up_blocks.append(up_block)
+            prev_output_channel = output_channel
+
+        # out
+        self.conv_norm_out = nn.GroupNorm(num_channels=block_out_channels[0], num_groups=norm_num_groups, eps=1e-05)
+        self.conv_act = get_activation("silu")
+        self.conv_out = nn.Conv2d(block_out_channels[0], out_channels, kernel_size=3, padding=1)
+
+    @property
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
+    def attn_processors(self) -> Dict[str, AttentionProcessor]:
+        r"""
+        Returns:
+            `dict` of attention processors: A dictionary containing all attention processors used in the model with
+            indexed by its weight name.
+        """
+        # set recursively
+        processors = {}
+
+        def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
+            if hasattr(module, "get_processor"):
+                processors[f"{name}.processor"] = module.get_processor(return_deprecated_lora=True)
+
+            for sub_name, child in module.named_children():
+                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
+
+            return processors
+
+        for name, module in self.named_children():
+            fn_recursive_add_processors(name, module, processors)
+
+        return processors
+
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
+    def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
+        r"""
+        Sets the attention processor to use to compute attention.
+
+        Parameters:
+            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
+                The instantiated processor class or a dictionary of processor classes that will be set as the processor
+                for **all** `Attention` layers.
+
+                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
+                processor. This is strongly recommended when setting trainable attention processors.
+
+        """
+        count = len(self.attn_processors.keys())
+
+        if isinstance(processor, dict) and len(processor) != count:
+            raise ValueError(
+                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
+                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
+            )
+
+        def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
+            if hasattr(module, "set_processor"):
+                if not isinstance(processor, dict):
+                    module.set_processor(processor)
+                else:
+                    module.set_processor(processor.pop(f"{name}.processor"))
+
+            for sub_name, child in module.named_children():
+                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
+
+        for name, module in self.named_children():
+            fn_recursive_attn_processor(name, module, processor)
+
+    # Copied from diffusers.models.unets.unet_3d_condition.UNet3DConditionModel.enable_forward_chunking
+    def enable_forward_chunking(self, chunk_size: Optional[int] = None, dim: int = 0) -> None:
+        """
+        Sets the attention processor to use [feed forward
+        chunking](https://huggingface.co/blog/reformer#2-chunked-feed-forward-layers).
+
+        Parameters:
+            chunk_size (`int`, *optional*):
+                The chunk size of the feed-forward layers. If not specified, will run feed-forward layer individually
+                over each tensor of dim=`dim`.
+            dim (`int`, *optional*, defaults to `0`):
+                The dimension over which the feed-forward computation should be chunked. Choose between dim=0 (batch)
+                or dim=1 (sequence length).
+        """
+        if dim not in [0, 1]:
+            raise ValueError(f"Make sure to set `dim` to either 0 or 1, not {dim}")
+
+        # By default chunk size is 1
+        chunk_size = chunk_size or 1
+
+        def fn_recursive_feed_forward(module: torch.nn.Module, chunk_size: int, dim: int):
+            if hasattr(module, "set_chunk_feed_forward"):
+                module.set_chunk_feed_forward(chunk_size=chunk_size, dim=dim)
+
+            for child in module.children():
+                fn_recursive_feed_forward(child, chunk_size, dim)
+
+        for module in self.children():
+            fn_recursive_feed_forward(module, chunk_size, dim)
+
+    # Copied from diffusers.models.unets.unet_3d_condition.UNet3DConditionModel.disable_forward_chunking
+    def disable_forward_chunking(self):
+        def fn_recursive_feed_forward(module: torch.nn.Module, chunk_size: int, dim: int):
+            if hasattr(module, "set_chunk_feed_forward"):
+                module.set_chunk_feed_forward(chunk_size=chunk_size, dim=dim)
+
+            for child in module.children():
+                fn_recursive_feed_forward(child, chunk_size, dim)
+
+        for module in self.children():
+            fn_recursive_feed_forward(module, None, 0)
+
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_default_attn_processor
+    def set_default_attn_processor(self):
+        """
+        Disables custom attention processors and sets the default attention implementation.
+        """
+        if all(proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+            processor = AttnAddedKVProcessor()
+        elif all(proc.__class__ in CROSS_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+            processor = AttnProcessor()
+        else:
+            raise ValueError(
+                f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
+            )
+
+        self.set_attn_processor(processor)
+
+    # Copied from diffusers.models.unets.unet_3d_condition.UNet3DConditionModel._set_gradient_checkpointing
+    def _set_gradient_checkpointing(self, module, value: bool = False) -> None:
+        if isinstance(module, (CrossAttnDownBlock3D, DownBlock3D, CrossAttnUpBlock3D, UpBlock3D)):
+            module.gradient_checkpointing = value
+
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.enable_freeu
+    def enable_freeu(self, s1, s2, b1, b2):
+        r"""Enables the FreeU mechanism from https://arxiv.org/abs/2309.11497.
+
+        The suffixes after the scaling factors represent the stage blocks where they are being applied.
+
+        Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of values that
+        are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL.
+
+        Args:
+            s1 (`float`):
+                Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to
+                mitigate the "oversmoothing effect" in the enhanced denoising process.
+            s2 (`float`):
+                Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to
+                mitigate the "oversmoothing effect" in the enhanced denoising process.
+            b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features.
+            b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features.
+        """
+        for i, upsample_block in enumerate(self.up_blocks):
+            setattr(upsample_block, "s1", s1)
+            setattr(upsample_block, "s2", s2)
+            setattr(upsample_block, "b1", b1)
+            setattr(upsample_block, "b2", b2)
+
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.disable_freeu
+    def disable_freeu(self):
+        """Disables the FreeU mechanism."""
+        freeu_keys = {"s1", "s2", "b1", "b2"}
+        for i, upsample_block in enumerate(self.up_blocks):
+            for k in freeu_keys:
+                if hasattr(upsample_block, k) or getattr(upsample_block, k, None) is not None:
+                    setattr(upsample_block, k, None)
+
+    def forward(
+        self,
+        sample: torch.FloatTensor,
+        timestep: Union[torch.Tensor, float, int],
+        fps: torch.Tensor,
+        image_latents: torch.Tensor,
+        image_embeddings: Optional[torch.Tensor] = None,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        timestep_cond: Optional[torch.Tensor] = None,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        return_dict: bool = True,
+    ) -> Union[UNet3DConditionOutput, Tuple[torch.FloatTensor]]:
+        r"""
+        The [`I2VGenXLUNet`] forward method.
+
+        Args:
+            sample (`torch.FloatTensor`):
+                The noisy input tensor with the following shape `(batch, num_frames, channel, height, width`.
+            timestep (`torch.FloatTensor` or `float` or `int`): The number of timesteps to denoise an input.
+            fps (`torch.Tensor`): Frames per second for the video being generated. Used as a "micro-condition".
+            image_latents (`torch.FloatTensor`): Image encodings from the VAE.
+            image_embeddings (`torch.FloatTensor`): Projection embeddings of the conditioning image computed with a vision encoder.
+            encoder_hidden_states (`torch.FloatTensor`):
+                The encoder hidden states with shape `(batch, sequence_length, feature_dim)`.
+            cross_attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~models.unet_3d_condition.UNet3DConditionOutput`] instead of a plain
+                tuple.
+
+        Returns:
+            [`~models.unet_3d_condition.UNet3DConditionOutput`] or `tuple`:
+                If `return_dict` is True, an [`~models.unet_3d_condition.UNet3DConditionOutput`] is returned, otherwise
+                a `tuple` is returned where the first element is the sample tensor.
+        """
+        batch_size, channels, num_frames, height, width = sample.shape
+
+        # By default samples have to be AT least a multiple of the overall upsampling factor.
+        # The overall upsampling factor is equal to 2 ** (# num of upsampling layears).
+        # However, the upsampling interpolation output size can be forced to fit any upsampling size
+        # on the fly if necessary.
+        default_overall_up_factor = 2**self.num_upsamplers
+
+        # upsample size should be forwarded when sample is not a multiple of `default_overall_up_factor`
+        forward_upsample_size = False
+        upsample_size = None
+
+        if any(s % default_overall_up_factor != 0 for s in sample.shape[-2:]):
+            logger.info("Forward upsample size to force interpolation output size.")
+            forward_upsample_size = True
+
+        # 1. time
+        timesteps = _to_tensor(timestep, sample.device)
+
+        # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+        timesteps = timesteps.expand(sample.shape[0])
+        t_emb = self.time_proj(timesteps)
+
+        # timesteps does not contain any weights and will always return f32 tensors
+        # but time_embedding might actually be running in fp16. so we need to cast here.
+        # there might be better ways to encapsulate this.
+        t_emb = t_emb.to(dtype=self.dtype)
+        t_emb = self.time_embedding(t_emb, timestep_cond)
+
+        # 2. FPS
+        # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+        fps = fps.expand(fps.shape[0])
+        fps_emb = self.fps_embedding(self.time_proj(fps).to(dtype=self.dtype))
+
+        # 3. time + FPS embeddings.
+        emb = t_emb + fps_emb
+        emb = emb.repeat_interleave(repeats=num_frames, dim=0)
+
+        # 4. context embeddings.
+        # The context embeddings consist of both text embeddings from the input prompt
+        # AND the image embeddings from the input image. For images, both VAE encodings
+        # and the CLIP image embeddings are incorporated.
+        # So the final `context_embeddings` becomes the query for cross-attention.
+        context_emb = sample.new_zeros(batch_size, 0, self.config.cross_attention_dim)
+        context_emb = torch.cat([context_emb, encoder_hidden_states], dim=1)
+
+        image_latents_context_embs = _collapse_frames_into_batch(image_latents[:, :, :1, :])
+        image_latents_context_embs = self.image_latents_context_embedding(image_latents_context_embs)
+
+        _batch_size, _channels, _height, _width = image_latents_context_embs.shape
+        image_latents_context_embs = image_latents_context_embs.permute(0, 2, 3, 1).reshape(
+            _batch_size, _height * _width, _channels
+        )
+        context_emb = torch.cat([context_emb, image_latents_context_embs], dim=1)
+
+        image_emb = self.context_embedding(image_embeddings)
+        image_emb = image_emb.view(-1, self.config.in_channels, self.config.cross_attention_dim)
+        context_emb = torch.cat([context_emb, image_emb], dim=1)
+        context_emb = context_emb.repeat_interleave(repeats=num_frames, dim=0)
+
+        image_latents = _collapse_frames_into_batch(image_latents)
+        image_latents = self.image_latents_proj_in(image_latents)
+        image_latents = (
+            image_latents[None, :]
+            .reshape(batch_size, num_frames, channels, height, width)
+            .permute(0, 3, 4, 1, 2)
+            .reshape(batch_size * height * width, num_frames, channels)
+        )
+        image_latents = self.image_latents_temporal_encoder(image_latents)
+        image_latents = image_latents.reshape(batch_size, height, width, num_frames, channels).permute(0, 4, 3, 1, 2)
+
+        # 5. pre-process
+        sample = torch.cat([sample, image_latents], dim=1)
+        sample = sample.permute(0, 2, 1, 3, 4).reshape((sample.shape[0] * num_frames, -1) + sample.shape[3:])
+        sample = self.conv_in(sample)
+        sample = self.transformer_in(
+            sample,
+            num_frames=num_frames,
+            cross_attention_kwargs=cross_attention_kwargs,
+            return_dict=False,
+        )[0]
+
+        # 6. down
+        down_block_res_samples = (sample,)
+        for downsample_block in self.down_blocks:
+            if hasattr(downsample_block, "has_cross_attention") and downsample_block.has_cross_attention:
+                sample, res_samples = downsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    encoder_hidden_states=context_emb,
+                    num_frames=num_frames,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                )
+            else:
+                sample, res_samples = downsample_block(hidden_states=sample, temb=emb, num_frames=num_frames)
+
+            down_block_res_samples += res_samples
+
+        # 7. mid
+        if self.mid_block is not None:
+            sample = self.mid_block(
+                sample,
+                emb,
+                encoder_hidden_states=context_emb,
+                num_frames=num_frames,
+                cross_attention_kwargs=cross_attention_kwargs,
+            )
+        # 8. up
+        for i, upsample_block in enumerate(self.up_blocks):
+            is_final_block = i == len(self.up_blocks) - 1
+
+            res_samples = down_block_res_samples[-len(upsample_block.resnets) :]
+            down_block_res_samples = down_block_res_samples[: -len(upsample_block.resnets)]
+
+            # if we have not reached the final block and need to forward the
+            # upsample size, we do it here
+            if not is_final_block and forward_upsample_size:
+                upsample_size = down_block_res_samples[-1].shape[2:]
+
+            if hasattr(upsample_block, "has_cross_attention") and upsample_block.has_cross_attention:
+                sample = upsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    res_hidden_states_tuple=res_samples,
+                    encoder_hidden_states=context_emb,
+                    upsample_size=upsample_size,
+                    num_frames=num_frames,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                )
+            else:
+                sample = upsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    res_hidden_states_tuple=res_samples,
+                    upsample_size=upsample_size,
+                    num_frames=num_frames,
+                )
+
+        # 9. post-process
+        sample = self.conv_norm_out(sample)
+        sample = self.conv_act(sample)
+
+        sample = self.conv_out(sample)
+
+        # reshape to (batch, channel, framerate, width, height)
+        sample = sample[None, :].reshape((-1, num_frames) + sample.shape[1:]).permute(0, 2, 1, 3, 4)
+
+        if not return_dict:
+            return (sample,)
+
+        return UNet3DConditionOutput(sample=sample)

src/diffusers/pipelines/__init__.py

@@ -220,6 +220,7 @@ else:
         "TextToVideoZeroSDXLPipeline",
         "VideoToVideoSDPipeline",
     ]
+    _import_structure["i2vgen_xl"] = ["I2VGenXLPipeline"]
     _import_structure["unclip"] = ["UnCLIPImageVariationPipeline", "UnCLIPPipeline"]
     _import_structure["unidiffuser"] = [
         "ImageTextPipelineOutput",
@@ -384,6 +385,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             VersatileDiffusionTextToImagePipeline,
             VQDiffusionPipeline,
         )
+        from .i2vgen_xl import I2VGenXLPipeline
         from .kandinsky import (
             KandinskyCombinedPipeline,
             KandinskyImg2ImgCombinedPipeline,

src/diffusers/pipelines/i2vgen_xl/__init__.py

+from typing import TYPE_CHECKING
+
+from ...utils import (
+    DIFFUSERS_SLOW_IMPORT,
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    get_objects_from_module,
+    is_torch_available,
+    is_transformers_available,
+)
+
+
+_dummy_objects = {}
+_import_structure = {}
+
+try:
+    if not (is_transformers_available() and is_torch_available()):
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    from ...utils import dummy_torch_and_transformers_objects  # noqa F403
+
+    _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
+else:
+    _import_structure["pipeline_i2vgen_xl"] = ["I2VGenXLPipeline"]
+
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+    try:
+        if not (is_transformers_available() and is_torch_available()):
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        from ...utils.dummy_torch_and_transformers_objects import *  # noqa F403
+    else:
+        from .pipeline_i2vgen_xl import I2VGenXLPipeline
+
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(
+        __name__,
+        globals()["__file__"],
+        _import_structure,
+        module_spec=__spec__,
+    )
+    for name, value in _dummy_objects.items():
+        setattr(sys.modules[__name__], name, value)

src/diffusers/pipelines/i2vgen_xl/pipeline_i2vgen_xl.py

+# Copyright 2023 Alibaba DAMO-VILAB and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+import numpy as np
+import PIL
+import torch
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection
+
+from ...image_processor import PipelineImageInput, VaeImageProcessor
+from ...loaders import LoraLoaderMixin
+from ...models import AutoencoderKL
+from ...models.lora import adjust_lora_scale_text_encoder
+from ...models.unets.unet_i2vgen_xl import I2VGenXLUNet
+from ...schedulers import DDIMScheduler
+from ...utils import (
+    USE_PEFT_BACKEND,
+    BaseOutput,
+    logging,
+    replace_example_docstring,
+    scale_lora_layers,
+    unscale_lora_layers,
+)
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+    Examples:
+        ```py
+        >>> import torch
+        >>> from diffusers import I2VGenXLPipeline
+
+        >>> pipeline = I2VGenXLPipeline.from_pretrained("ali-vilab/i2vgen-xl", torch_dtype=torch.float16, variant="fp16")
+        >>> pipeline.enable_model_cpu_offload()
+
+        >>> image_url = "https://github.com/ali-vilab/i2vgen-xl/blob/main/data/test_images/img_0009.png?raw=true"
+        >>> image = load_image(image_url).convert("RGB")
+
+        >>> prompt = "Papers were floating in the air on a table in the library"
+        >>> negative_prompt = "Distorted, discontinuous, Ugly, blurry, low resolution, motionless, static, disfigured, disconnected limbs, Ugly faces, incomplete arms"
+        >>> generator = torch.manual_seed(8888)
+
+        >>> frames = pipeline(
+        ...     prompt=prompt,
+        ...     image=image,
+        ...     num_inference_steps=50,
+        ...     negative_prompt=negative_prompt,
+        ...     guidance_scale=9.0,
+        ...     generator=generator
+        ... ).frames[0]
+        >>> video_path = export_to_gif(frames, "i2v.gif")
+        ```
+"""
+
+
+# Copied from diffusers.pipelines.animatediff.pipeline_animatediff.tensor2vid
+def tensor2vid(video: torch.Tensor, processor: "VaeImageProcessor", output_type: str = "np"):
+    batch_size, channels, num_frames, height, width = video.shape
+    outputs = []
+    for batch_idx in range(batch_size):
+        batch_vid = video[batch_idx].permute(1, 0, 2, 3)
+        batch_output = processor.postprocess(batch_vid, output_type)
+
+        outputs.append(batch_output)
+
+    if output_type == "np":
+        outputs = np.stack(outputs)
+
+    elif output_type == "pt":
+        outputs = torch.stack(outputs)
+
+    elif not output_type == "pil":
+        raise ValueError(f"{output_type} does not exist. Please choose one of ['np', 'pt', 'pil]")
+
+    return outputs
+
+
+@dataclass
+class I2VGenXLPipelineOutput(BaseOutput):
+    r"""
+    Output class for image-to-video pipeline.
+
+    Args:
+        frames (`List[np.ndarray]` or `torch.FloatTensor`)
+            List of denoised frames (essentially images) as NumPy arrays of shape `(height, width, num_channels)` or as
+            a `torch` tensor. The length of the list denotes the video length (the number of frames).
+    """
+
+    frames: Union[List[np.ndarray], torch.FloatTensor]
+
+
+class I2VGenXLPipeline(DiffusionPipeline):
+    r"""
+    Pipeline for image-to-video generation as proposed in [I2VGenXL](https://i2vgen-xl.github.io/).
+
+    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+    implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+    Args:
+        vae ([`AutoencoderKL`]):
+            Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+        text_encoder ([`CLIPTextModel`]):
+            Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
+        tokenizer (`CLIPTokenizer`):
+            A [`~transformers.CLIPTokenizer`] to tokenize text.
+        unet ([`I2VGenXLUNet`]):
+            A [`I2VGenXLUNet`] to denoise the encoded video latents.
+        scheduler ([`DDIMScheduler`]):
+            A scheduler to be used in combination with `unet` to denoise the encoded image latents.
+    """
+
+    model_cpu_offload_seq = "text_encoder->image_encoder->unet->vae"
+
+    def __init__(
+        self,
+        vae: AutoencoderKL,
+        text_encoder: CLIPTextModel,
+        tokenizer: CLIPTokenizer,
+        image_encoder: CLIPVisionModelWithProjection,
+        feature_extractor: CLIPImageProcessor,
+        unet: I2VGenXLUNet,
+        scheduler: DDIMScheduler,
+    ):
+        super().__init__()
+
+        self.register_modules(
+            vae=vae,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            image_encoder=image_encoder,
+            feature_extractor=feature_extractor,
+            unet=unet,
+            scheduler=scheduler,
+        )
+        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+        # `do_resize=False` as we do custom resizing.
+        self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor, do_resize=False)
+
+    @property
+    def guidance_scale(self):
+        return self._guidance_scale
+
+    # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+    # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+    # corresponds to doing no classifier free guidance.
+    @property
+    def do_classifier_free_guidance(self):
+        return self._guidance_scale > 1
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_slicing
+    def enable_vae_slicing(self):
+        r"""
+        Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
+        compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
+        """
+        self.vae.enable_slicing()
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_slicing
+    def disable_vae_slicing(self):
+        r"""
+        Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_slicing()
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_tiling
+    def enable_vae_tiling(self):
+        r"""
+        Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
+        compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
+        processing larger images.
+        """
+        self.vae.enable_tiling()
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_tiling
+    def disable_vae_tiling(self):
+        r"""
+        Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_tiling()
+
+    def encode_prompt(
+        self,
+        prompt,
+        device,
+        num_videos_per_prompt,
+        negative_prompt=None,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        lora_scale: Optional[float] = None,
+        clip_skip: Optional[int] = None,
+    ):
+        r"""
+        Encodes the prompt into text encoder hidden states.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                prompt to be encoded
+            device: (`torch.device`):
+                torch device
+            num_videos_per_prompt (`int`):
+                number of images that should be generated per prompt
+            do_classifier_free_guidance (`bool`):
+                whether to use classifier free guidance or not
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            lora_scale (`float`, *optional*):
+                A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+            clip_skip (`int`, *optional*):
+                Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+                the output of the pre-final layer will be used for computing the prompt embeddings.
+        """
+        # set lora scale so that monkey patched LoRA
+        # function of text encoder can correctly access it
+        if lora_scale is not None and isinstance(self, LoraLoaderMixin):
+            self._lora_scale = lora_scale
+
+            # dynamically adjust the LoRA scale
+            if not USE_PEFT_BACKEND:
+                adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
+            else:
+                scale_lora_layers(self.text_encoder, lora_scale)
+
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        if prompt_embeds is None:
+            text_inputs = self.tokenizer(
+                prompt,
+                padding="max_length",
+                max_length=self.tokenizer.model_max_length,
+                truncation=True,
+                return_tensors="pt",
+            )
+            text_input_ids = text_inputs.input_ids
+            untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+            if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+                text_input_ids, untruncated_ids
+            ):
+                removed_text = self.tokenizer.batch_decode(
+                    untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
+                )
+                logger.warning(
+                    "The following part of your input was truncated because CLIP can only handle sequences up to"
+                    f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+                )
+
+            if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+                attention_mask = text_inputs.attention_mask.to(device)
+            else:
+                attention_mask = None
+
+            if clip_skip is None:
+                prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask)
+                prompt_embeds = prompt_embeds[0]
+            else:
+                prompt_embeds = self.text_encoder(
+                    text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True
+                )
+                # Access the `hidden_states` first, that contains a tuple of
+                # all the hidden states from the encoder layers. Then index into
+                # the tuple to access the hidden states from the desired layer.
+                prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)]
+                # We also need to apply the final LayerNorm here to not mess with the
+                # representations. The `last_hidden_states` that we typically use for
+                # obtaining the final prompt representations passes through the LayerNorm
+                # layer.
+                prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds)
+
+        if self.text_encoder is not None:
+            prompt_embeds_dtype = self.text_encoder.dtype
+        elif self.unet is not None:
+            prompt_embeds_dtype = self.unet.dtype
+        else:
+            prompt_embeds_dtype = prompt_embeds.dtype
+
+        prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+        bs_embed, seq_len, _ = prompt_embeds.shape
+        # duplicate text embeddings for each generation per prompt, using mps friendly method
+        prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1)
+        prompt_embeds = prompt_embeds.view(bs_embed * num_videos_per_prompt, seq_len, -1)
+
+        # get unconditional embeddings for classifier free guidance
+        if self.do_classifier_free_guidance and negative_prompt_embeds is None:
+            uncond_tokens: List[str]
+            if negative_prompt is None:
+                uncond_tokens = [""] * batch_size
+            elif prompt is not None and type(prompt) is not type(negative_prompt):
+                raise TypeError(
+                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+                    f" {type(prompt)}."
+                )
+            elif isinstance(negative_prompt, str):
+                uncond_tokens = [negative_prompt]
+            elif batch_size != len(negative_prompt):
+                raise ValueError(
+                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+                    " the batch size of `prompt`."
+                )
+            else:
+                uncond_tokens = negative_prompt
+
+            max_length = prompt_embeds.shape[1]
+            uncond_input = self.tokenizer(
+                uncond_tokens,
+                padding="max_length",
+                max_length=max_length,
+                truncation=True,
+                return_tensors="pt",
+            )
+
+            if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+                attention_mask = uncond_input.attention_mask.to(device)
+            else:
+                attention_mask = None
+
+            # Apply clip_skip to negative prompt embeds
+            if clip_skip is None:
+                negative_prompt_embeds = self.text_encoder(
+                    uncond_input.input_ids.to(device),
+                    attention_mask=attention_mask,
+                )
+                negative_prompt_embeds = negative_prompt_embeds[0]
+            else:
+                negative_prompt_embeds = self.text_encoder(
+                    uncond_input.input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True
+                )
+                # Access the `hidden_states` first, that contains a tuple of
+                # all the hidden states from the encoder layers. Then index into
+                # the tuple to access the hidden states from the desired layer.
+                negative_prompt_embeds = negative_prompt_embeds[-1][-(clip_skip + 1)]
+                # We also need to apply the final LayerNorm here to not mess with the
+                # representations. The `last_hidden_states` that we typically use for
+                # obtaining the final prompt representations passes through the LayerNorm
+                # layer.
+                negative_prompt_embeds = self.text_encoder.text_model.final_layer_norm(negative_prompt_embeds)
+
+        if self.do_classifier_free_guidance:
+            # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+            seq_len = negative_prompt_embeds.shape[1]
+
+            negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+            negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_videos_per_prompt, 1)
+            negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1)
+
+        if isinstance(self, LoraLoaderMixin) and USE_PEFT_BACKEND:
+            # Retrieve the original scale by scaling back the LoRA layers
+            unscale_lora_layers(self.text_encoder, lora_scale)
+
+        return prompt_embeds, negative_prompt_embeds
+
+    def _encode_image(self, image, device, num_videos_per_prompt):
+        dtype = next(self.image_encoder.parameters()).dtype
+
+        if not isinstance(image, torch.Tensor):
+            image = self.image_processor.pil_to_numpy(image)
+            image = self.image_processor.numpy_to_pt(image)
+
+            # Normalize the image with CLIP training stats.
+            image = self.feature_extractor(
+                images=image,
+                do_normalize=True,
+                do_center_crop=False,
+                do_resize=False,
+                do_rescale=False,
+                return_tensors="pt",
+            ).pixel_values
+
+        image = image.to(device=device, dtype=dtype)
+        image_embeddings = self.image_encoder(image).image_embeds
+        image_embeddings = image_embeddings.unsqueeze(1)
+
+        # duplicate image embeddings for each generation per prompt, using mps friendly method
+        bs_embed, seq_len, _ = image_embeddings.shape
+        image_embeddings = image_embeddings.repeat(1, num_videos_per_prompt, 1)
+        image_embeddings = image_embeddings.view(bs_embed * num_videos_per_prompt, seq_len, -1)
+
+        if self.do_classifier_free_guidance:
+            negative_image_embeddings = torch.zeros_like(image_embeddings)
+            image_embeddings = torch.cat([negative_image_embeddings, image_embeddings])
+
+        return image_embeddings
+
+    def decode_latents(self, latents, decode_chunk_size=None):
+        latents = 1 / self.vae.config.scaling_factor * latents
+
+        batch_size, channels, num_frames, height, width = latents.shape
+        latents = latents.permute(0, 2, 1, 3, 4).reshape(batch_size * num_frames, channels, height, width)
+
+        if decode_chunk_size is not None:
+            frames = []
+            for i in range(0, latents.shape[0], decode_chunk_size):
+                frame = self.vae.decode(latents[i : i + decode_chunk_size]).sample
+                frames.append(frame)
+            image = torch.cat(frames, dim=0)
+        else:
+            image = self.vae.decode(latents).sample
+
+        decode_shape = (batch_size, num_frames, -1) + image.shape[2:]
+        video = image[None, :].reshape(decode_shape).permute(0, 2, 1, 3, 4)
+
+        # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
+        video = video.float()
+        return video
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+    def prepare_extra_step_kwargs(self, generator, eta):
+        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+        # and should be between [0, 1]
+
+        accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        extra_step_kwargs = {}
+        if accepts_eta:
+            extra_step_kwargs["eta"] = eta
+
+        # check if the scheduler accepts generator
+        accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        if accepts_generator:
+            extra_step_kwargs["generator"] = generator
+        return extra_step_kwargs
+
+    def check_inputs(
+        self,
+        prompt,
+        image,
+        height,
+        width,
+        negative_prompt=None,
+        prompt_embeds=None,
+        negative_prompt_embeds=None,
+    ):
+        if height % 8 != 0 or width % 8 != 0:
+            raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+        if prompt is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt is None and prompt_embeds is None:
+            raise ValueError(
+                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+            )
+        elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+        if negative_prompt is not None and negative_prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+            )
+
+        if prompt_embeds is not None and negative_prompt_embeds is not None:
+            if prompt_embeds.shape != negative_prompt_embeds.shape:
+                raise ValueError(
+                    "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+                    f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+                    f" {negative_prompt_embeds.shape}."
+                )
+
+        if (
+            not isinstance(image, torch.Tensor)
+            and not isinstance(image, PIL.Image.Image)
+            and not isinstance(image, list)
+        ):
+            raise ValueError(
+                "`image` has to be of type `torch.FloatTensor` or `PIL.Image.Image` or `List[PIL.Image.Image]` but is"
+                f" {type(image)}"
+            )
+
+    def prepare_image_latents(
+        self,
+        image,
+        device,
+        num_frames,
+        num_videos_per_prompt,
+    ):
+        image = image.to(device=device)
+        image_latents = self.vae.encode(image).latent_dist.sample()
+        image_latents = image_latents * self.vae.config.scaling_factor
+
+        # Add frames dimension to image latents
+        image_latents = image_latents.unsqueeze(2)
+
+        # Append a position mask for each subsequent frame
+        # after the intial image latent frame
+        frame_position_mask = []
+        for frame_idx in range(num_frames - 1):
+            scale = (frame_idx + 1) / (num_frames - 1)
+            frame_position_mask.append(torch.ones_like(image_latents[:, :, :1]) * scale)
+        if frame_position_mask:
+            frame_position_mask = torch.cat(frame_position_mask, dim=2)
+            image_latents = torch.cat([image_latents, frame_position_mask], dim=2)
+
+        # duplicate image_latents for each generation per prompt, using mps friendly method
+        image_latents = image_latents.repeat(num_videos_per_prompt, 1, 1, 1, 1)
+
+        if self.do_classifier_free_guidance:
+            image_latents = torch.cat([image_latents] * 2)
+
+        return image_latents
+
+    # Copied from diffusers.pipelines.text_to_video_synthesis.pipeline_text_to_video_synth.TextToVideoSDPipeline.prepare_latents
+    def prepare_latents(
+        self, batch_size, num_channels_latents, num_frames, height, width, dtype, device, generator, latents=None
+    ):
+        shape = (
+            batch_size,
+            num_channels_latents,
+            num_frames,
+            height // self.vae_scale_factor,
+            width // self.vae_scale_factor,
+        )
+        if isinstance(generator, list) and len(generator) != batch_size:
+            raise ValueError(
+                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+            )
+
+        if latents is None:
+            latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+        else:
+            latents = latents.to(device)
+
+        # scale the initial noise by the standard deviation required by the scheduler
+        latents = latents * self.scheduler.init_noise_sigma
+        return latents
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_freeu
+    def enable_freeu(self, s1: float, s2: float, b1: float, b2: float):
+        r"""Enables the FreeU mechanism as in https://arxiv.org/abs/2309.11497.
+
+        The suffixes after the scaling factors represent the stages where they are being applied.
+
+        Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of the values
+        that are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL.
+
+        Args:
+            s1 (`float`):
+                Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to
+                mitigate "oversmoothing effect" in the enhanced denoising process.
+            s2 (`float`):
+                Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to
+                mitigate "oversmoothing effect" in the enhanced denoising process.
+            b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features.
+            b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features.
+        """
+        if not hasattr(self, "unet"):
+            raise ValueError("The pipeline must have `unet` for using FreeU.")
+        self.unet.enable_freeu(s1=s1, s2=s2, b1=b1, b2=b2)
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_freeu
+    def disable_freeu(self):
+        """Disables the FreeU mechanism if enabled."""
+        self.unet.disable_freeu()
+
+    @torch.no_grad()
+    @replace_example_docstring(EXAMPLE_DOC_STRING)
+    def __call__(
+        self,
+        prompt: Union[str, List[str]] = None,
+        image: PipelineImageInput = None,
+        height: Optional[int] = 704,
+        width: Optional[int] = 1280,
+        target_fps: Optional[int] = 16,
+        num_frames: int = 16,
+        num_inference_steps: int = 50,
+        guidance_scale: float = 9.0,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        eta: float = 0.0,
+        num_videos_per_prompt: Optional[int] = 1,
+        decode_chunk_size: Optional[int] = 1,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        clip_skip: Optional[int] = 1,
+    ):
+        r"""
+        The call function to the pipeline for image-to-video generation with [`I2VGenXLPipeline`].
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
+            image (`PIL.Image.Image` or `List[PIL.Image.Image]` or `torch.FloatTensor`):
+                Image or images to guide image generation. If you provide a tensor, it needs to be compatible with
+                [`CLIPImageProcessor`](https://huggingface.co/lambdalabs/sd-image-variations-diffusers/blob/main/feature_extractor/preprocessor_config.json).
+            height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+                The height in pixels of the generated image.
+            width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+                The width in pixels of the generated image.
+            target_fps (`int`, *optional*):
+                Frames per second. The rate at which the generated images shall be exported to a video after generation. This is also used as a "micro-condition" while generation.
+            num_frames (`int`, *optional*):
+                The number of video frames to generate.
+            num_inference_steps (`int`, *optional*):
+                The number of denoising steps.
+            guidance_scale (`float`, *optional*, defaults to 7.5):
+                A higher guidance scale value encourages the model to generate images closely linked to the text
+                `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide what to not include in image generation. If not defined, you need to
+                pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
+            eta (`float`, *optional*):
+                Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+                to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+            num_videos_per_prompt (`int`, *optional*):
+                The number of images to generate per prompt.
+            decode_chunk_size (`int`, *optional*):
+                The number of frames to decode at a time. The higher the chunk size, the higher the temporal consistency
+                between frames, but also the higher the memory consumption. By default, the decoder will decode all frames at once
+                for maximal quality. Reduce `decode_chunk_size` to reduce memory usage.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+                generation deterministic.
+            latents (`torch.FloatTensor`, *optional*):
+                Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor is generated by sampling using the supplied random `generator`.
+            prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+                provided, text embeddings are generated from the `prompt` input argument.
+            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
+                not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+                plain tuple.
+            cross_attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
+                [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            clip_skip (`int`, *optional*):
+                Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+                the output of the pre-final layer will be used for computing the prompt embeddings.
+
+        Examples:
+
+        Returns:
+            [`pipelines.i2vgen_xl.pipeline_i2vgen_xl.I2VGenXLPipelineOutput`] or `tuple`:
+                If `return_dict` is `True`, [`pipelines.i2vgen_xl.pipeline_i2vgen_xl.I2VGenXLPipelineOutput`] is
+                returned, otherwise a `tuple` is returned where the first element is a list with the generated frames.
+        """
+        # 0. Default height and width to unet
+        height = height or self.unet.config.sample_size * self.vae_scale_factor
+        width = width or self.unet.config.sample_size * self.vae_scale_factor
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(prompt, image, height, width, negative_prompt, prompt_embeds, negative_prompt_embeds)
+
+        # 2. Define call parameters
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        device = self._execution_device
+        # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+        # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+        # corresponds to doing no classifier free guidance.
+        self._guidance_scale = guidance_scale
+
+        # 3.1 Encode input text prompt
+        text_encoder_lora_scale = (
+            cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None
+        )
+        prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+            prompt,
+            device,
+            num_videos_per_prompt,
+            negative_prompt,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            lora_scale=text_encoder_lora_scale,
+            clip_skip=clip_skip,
+        )
+        # For classifier free guidance, we need to do two forward passes.
+        # Here we concatenate the unconditional and text embeddings into a single batch
+        # to avoid doing two forward passes
+        if self.do_classifier_free_guidance:
+            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+        # 3.2 Encode image prompt
+        # 3.2.1 Image encodings.
+        # https://github.com/ali-vilab/i2vgen-xl/blob/2539c9262ff8a2a22fa9daecbfd13f0a2dbc32d0/tools/inferences/inference_i2vgen_entrance.py#L114
+        cropped_image = _center_crop_wide(image, (width, width))
+        cropped_image = _resize_bilinear(
+            cropped_image, (self.feature_extractor.crop_size["width"], self.feature_extractor.crop_size["height"])
+        )
+        image_embeddings = self._encode_image(cropped_image, device, num_videos_per_prompt)
+
+        # 3.2.2 Image latents.
+        resized_image = _center_crop_wide(image, (width, height))
+        image = self.image_processor.preprocess(resized_image).to(device=device, dtype=image_embeddings.dtype)
+        image_latents = self.prepare_image_latents(
+            image,
+            device=device,
+            num_frames=num_frames,
+            num_videos_per_prompt=num_videos_per_prompt,
+        )
+
+        # 3.3 Prepare additional conditions for the UNet.
+        if self.do_classifier_free_guidance:
+            fps_tensor = torch.tensor([target_fps, target_fps]).to(device)
+        else:
+            fps_tensor = torch.tensor([target_fps]).to(device)
+        fps_tensor = fps_tensor.repeat(batch_size * num_videos_per_prompt, 1).ravel()
+
+        # 4. Prepare timesteps
+        self.scheduler.set_timesteps(num_inference_steps, device=device)
+        timesteps = self.scheduler.timesteps
+
+        # 5. Prepare latent variables
+        num_channels_latents = self.unet.config.in_channels
+        latents = self.prepare_latents(
+            batch_size * num_videos_per_prompt,
+            num_channels_latents,
+            num_frames,
+            height,
+            width,
+            prompt_embeds.dtype,
+            device,
+            generator,
+            latents,
+        )
+
+        # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+        extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+        # 7. Denoising loop
+        num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                # expand the latents if we are doing classifier free guidance
+                latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
+                latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+                # predict the noise residual
+                noise_pred = self.unet(
+                    latent_model_input,
+                    t,
+                    encoder_hidden_states=prompt_embeds,
+                    fps=fps_tensor,
+                    image_latents=image_latents,
+                    image_embeddings=image_embeddings,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    return_dict=False,
+                )[0]
+
+                # perform guidance
+                if self.do_classifier_free_guidance:
+                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                    noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+                # reshape latents
+                batch_size, channel, frames, width, height = latents.shape
+                latents = latents.permute(0, 2, 1, 3, 4).reshape(batch_size * frames, channel, width, height)
+                noise_pred = noise_pred.permute(0, 2, 1, 3, 4).reshape(batch_size * frames, channel, width, height)
+
+                # compute the previous noisy sample x_t -> x_t-1
+                latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
+
+                # reshape latents back
+                latents = latents[None, :].reshape(batch_size, frames, channel, width, height).permute(0, 2, 1, 3, 4)
+                # call the callback, if provided
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                    progress_bar.update()
+
+        if output_type == "latent":
+            return I2VGenXLPipelineOutput(frames=latents)
+
+        video_tensor = self.decode_latents(latents, decode_chunk_size=decode_chunk_size)
+        video = tensor2vid(video_tensor, self.image_processor, output_type=output_type)
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return (video,)
+
+        return I2VGenXLPipelineOutput(frames=video)
+
+
+# The following utilities are taken and adapted from
+# https://github.com/ali-vilab/i2vgen-xl/blob/main/utils/transforms.py.
+
+
+def _convert_pt_to_pil(image: Union[torch.Tensor, List[torch.Tensor]]):
+    if isinstance(image, list) and isinstance(image[0], torch.Tensor):
+        image = torch.cat(image, 0)
+
+    if isinstance(image, torch.Tensor):
+        if image.ndim == 3:
+            image = image.unsqueeze(0)
+
+        image_numpy = VaeImageProcessor.pt_to_numpy(image)
+        image_pil = VaeImageProcessor.numpy_to_pil(image_numpy)
+        image = image_pil
+
+    return image
+
+
+def _resize_bilinear(
+    image: Union[torch.Tensor, List[torch.Tensor], PIL.Image.Image, List[PIL.Image.Image]], resolution: Tuple[int, int]
+):
+    # First convert the images to PIL in case they are float tensors (only relevant for tests now).
+    image = _convert_pt_to_pil(image)
+
+    if isinstance(image, list):
+        image = [u.resize(resolution, PIL.Image.BILINEAR) for u in image]
+    else:
+        image = image.resize(resolution, PIL.Image.BILINEAR)
+    return image
+
+
+def _center_crop_wide(
+    image: Union[torch.Tensor, List[torch.Tensor], PIL.Image.Image, List[PIL.Image.Image]], resolution: Tuple[int, int]
+):
+    # First convert the images to PIL in case they are float tensors (only relevant for tests now).
+    image = _convert_pt_to_pil(image)
+
+    if isinstance(image, list):
+        scale = min(image[0].size[0] / resolution[0], image[0].size[1] / resolution[1])
+        image = [u.resize((round(u.width // scale), round(u.height // scale)), resample=PIL.Image.BOX) for u in image]
+
+        # center crop
+        x1 = (image[0].width - resolution[0]) // 2
+        y1 = (image[0].height - resolution[1]) // 2
+        image = [u.crop((x1, y1, x1 + resolution[0], y1 + resolution[1])) for u in image]
+        return image
+    else:
+        scale = min(image.size[0] / resolution[0], image.size[1] / resolution[1])
+        image = image.resize((round(image.width // scale), round(image.height // scale)), resample=PIL.Image.BOX)
+        x1 = (image.width - resolution[0]) // 2
+        y1 = (image.height - resolution[1]) // 2
+        image = image.crop((x1, y1, x1 + resolution[0], y1 + resolution[1]))
+        return image

src/diffusers/utils/dummy_pt_objects.py

@@ -92,6 +92,21 @@ class ControlNetModel(metaclass=DummyObject):
         requires_backends(cls, ["torch"])
 
 
+class I2VGenXLUNet(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+
 class Kandinsky3UNet(metaclass=DummyObject):
     _backends = ["torch"]
 

src/diffusers/utils/dummy_torch_and_transformers_objects.py

@@ -197,6 +197,21 @@ class CycleDiffusionPipeline(metaclass=DummyObject):
         requires_backends(cls, ["torch", "transformers"])
 
 
+class I2VGenXLPipeline(metaclass=DummyObject):
+    _backends = ["torch", "transformers"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch", "transformers"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+
 class IFImg2ImgPipeline(metaclass=DummyObject):
     _backends = ["torch", "transformers"]
 

tests/pipelines/i2vgen_xl/test_i2vgenxl.py

+# coding=utf-8
+# Copyright 2023 HuggingFace Inc.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import gc
+import random
+import unittest
+
+import numpy as np
+import torch
+from transformers import (
+    CLIPImageProcessor,
+    CLIPTextConfig,
+    CLIPTextModel,
+    CLIPTokenizer,
+    CLIPVisionConfig,
+    CLIPVisionModelWithProjection,
+)
+
+from diffusers import (
+    AutoencoderKL,
+    DDIMScheduler,
+    I2VGenXLPipeline,
+)
+from diffusers.models.unets import I2VGenXLUNet
+from diffusers.utils import is_xformers_available, load_image
+from diffusers.utils.testing_utils import (
+    enable_full_determinism,
+    floats_tensor,
+    numpy_cosine_similarity_distance,
+    print_tensor_test,
+    require_torch_gpu,
+    skip_mps,
+    slow,
+    torch_device,
+)
+
+from ..test_pipelines_common import PipelineTesterMixin
+
+
+enable_full_determinism()
+
+
+@skip_mps
+class I2VGenXLPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
+    pipeline_class = I2VGenXLPipeline
+    params = frozenset(["prompt", "negative_prompt", "image"])
+    batch_params = frozenset(["prompt", "negative_prompt", "image", "generator"])
+    # No `output_type`.
+    required_optional_params = frozenset(["num_inference_steps", "generator", "latents", "return_dict"])
+
+    def get_dummy_components(self):
+        torch.manual_seed(0)
+        scheduler = DDIMScheduler(
+            beta_start=0.00085,
+            beta_end=0.012,
+            beta_schedule="scaled_linear",
+            clip_sample=False,
+            set_alpha_to_one=False,
+        )
+
+        torch.manual_seed(0)
+        unet = I2VGenXLUNet(
+            block_out_channels=(4, 8),
+            layers_per_block=1,
+            sample_size=32,
+            in_channels=4,
+            out_channels=4,
+            down_block_types=("CrossAttnDownBlock3D", "DownBlock3D"),
+            up_block_types=("UpBlock3D", "CrossAttnUpBlock3D"),
+            cross_attention_dim=4,
+            num_attention_heads=4,
+            norm_num_groups=2,
+        )
+
+        torch.manual_seed(0)
+        vae = AutoencoderKL(
+            block_out_channels=(8,),
+            in_channels=3,
+            out_channels=3,
+            down_block_types=["DownEncoderBlock2D"],
+            up_block_types=["UpDecoderBlock2D"],
+            latent_channels=4,
+            sample_size=32,
+            norm_num_groups=2,
+        )
+        torch.manual_seed(0)
+        text_encoder_config = CLIPTextConfig(
+            bos_token_id=0,
+            eos_token_id=2,
+            hidden_size=4,
+            intermediate_size=16,
+            layer_norm_eps=1e-05,
+            num_attention_heads=2,
+            num_hidden_layers=2,
+            pad_token_id=1,
+            vocab_size=1000,
+            hidden_act="gelu",
+            projection_dim=32,
+        )
+        text_encoder = CLIPTextModel(text_encoder_config)
+        tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
+
+        torch.manual_seed(0)
+        vision_encoder_config = CLIPVisionConfig(
+            hidden_size=4,
+            projection_dim=4,
+            num_hidden_layers=2,
+            num_attention_heads=2,
+            image_size=32,
+            intermediate_size=16,
+            patch_size=1,
+        )
+        image_encoder = CLIPVisionModelWithProjection(vision_encoder_config)
+
+        torch.manual_seed(0)
+        feature_extractor = CLIPImageProcessor(crop_size=32, size=32)
+
+        components = {
+            "unet": unet,
+            "scheduler": scheduler,
+            "vae": vae,
+            "text_encoder": text_encoder,
+            "image_encoder": image_encoder,
+            "tokenizer": tokenizer,
+            "feature_extractor": feature_extractor,
+        }
+        return components
+
+    def get_dummy_inputs(self, device, seed=0):
+        if str(device).startswith("mps"):
+            generator = torch.manual_seed(seed)
+        else:
+            generator = torch.Generator(device=device).manual_seed(seed)
+
+        input_image = floats_tensor((1, 3, 32, 32), rng=random.Random(seed)).to(device)
+        inputs = {
+            "prompt": "A painting of a squirrel eating a burger",
+            "image": input_image,
+            "generator": generator,
+            "num_inference_steps": 2,
+            "guidance_scale": 6.0,
+            "output_type": "pt",
+            "num_frames": 4,
+            "width": 32,
+            "height": 32,
+        }
+        return inputs
+
+    def test_text_to_video_default_case(self):
+        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+        pipe = pipe.to(device)
+        pipe.set_progress_bar_config(disable=None)
+
+        inputs = self.get_dummy_inputs(device)
+        inputs["output_type"] = "np"
+        frames = pipe(**inputs).frames
+
+        image_slice = frames[0][0][-3:, -3:, -1]
+
+        assert frames[0][0].shape == (32, 32, 3)
+        expected_slice = np.array([0.5146, 0.6525, 0.6032, 0.5204, 0.5675, 0.4125, 0.3016, 0.5172, 0.4095])
+
+        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2
+
+    def test_save_load_local(self):
+        super().test_save_load_local(expected_max_difference=0.006)
+
+    def test_sequential_cpu_offload_forward_pass(self):
+        super().test_sequential_cpu_offload_forward_pass(expected_max_diff=0.008)
+
+    def test_dict_tuple_outputs_equivalent(self):
+        super().test_dict_tuple_outputs_equivalent(expected_max_difference=0.008)
+
+    def test_save_load_optional_components(self):
+        super().test_save_load_optional_components(expected_max_difference=0.008)
+
+    @unittest.skip("Deprecated functionality")
+    def test_attention_slicing_forward_pass(self):
+        pass
+
+    @unittest.skipIf(
+        torch_device != "cuda" or not is_xformers_available(),
+        reason="XFormers attention is only available with CUDA and `xformers` installed",
+    )
+    def test_xformers_attention_forwardGenerator_pass(self):
+        self._test_xformers_attention_forwardGenerator_pass(test_mean_pixel_difference=False, expected_max_diff=1e-2)
+
+    def test_inference_batch_single_identical(self):
+        super().test_inference_batch_single_identical(batch_size=2, expected_max_diff=0.008)
+
+    def test_model_cpu_offload_forward_pass(self):
+        super().test_model_cpu_offload_forward_pass(expected_max_diff=0.008)
+
+    def test_num_videos_per_prompt(self):
+        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+        pipe = pipe.to(device)
+        pipe.set_progress_bar_config(disable=None)
+
+        inputs = self.get_dummy_inputs(device)
+        inputs["output_type"] = "np"
+        frames = pipe(**inputs, num_videos_per_prompt=2).frames
+
+        assert frames.shape == (2, 4, 32, 32, 3)
+        assert frames[0][0].shape == (32, 32, 3)
+
+        image_slice = frames[0][0][-3:, -3:, -1]
+        expected_slice = np.array([0.5146, 0.6525, 0.6032, 0.5204, 0.5675, 0.4125, 0.3016, 0.5172, 0.4095])
+
+        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2
+
+
+@slow
+@require_torch_gpu
+class I2VGenXLPipelineSlowTests(unittest.TestCase):
+    def tearDown(self):
+        # clean up the VRAM after each test
+        super().tearDown()
+        gc.collect()
+        torch.cuda.empty_cache()
+
+    def test_i2vgen_xl(self):
+        pipe = I2VGenXLPipeline.from_pretrained("ali-vilab/i2vgen-xl", torch_dtype=torch.float16, variant="fp16")
+        pipe = pipe.to(torch_device)
+        pipe.enable_model_cpu_offload()
+        pipe.set_progress_bar_config(disable=None)
+        image = load_image(
+            "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/pix2pix/cat_6.png?download=true"
+        )
+
+        generator = torch.Generator("cpu").manual_seed(0)
+        num_frames = 3
+
+        output = pipe(
+            image=image,
+            prompt="my cat",
+            num_frames=num_frames,
+            generator=generator,
+            num_inference_steps=3,
+            output_type="np",
+        )
+
+        image = output.frames[0]
+        assert image.shape == (num_frames, 704, 1280, 3)
+
+        image_slice = image[0, -3:, -3:, -1]
+        print_tensor_test(image_slice.flatten())
+        expected_slice = np.array([0.5482, 0.6244, 0.6274, 0.4584, 0.5935, 0.5937, 0.4579, 0.5767, 0.5892])
+        assert numpy_cosine_similarity_distance(image_slice.flatten(), expected_slice.flatten()) < 1e-3