Latte: Latent Diffusion Transformer for Video Generation (#8404)

* add Latte to diffusers

* remove print

* remove print

* remove print

* remove unuse codes

* remove layer_norm_latte and add a flag

* remove layer_norm_latte and add a flag

* update latte_pipeline

* update latte_pipeline

* remove unuse squeeze

* add norm_hidden_states.ndim == 2: # for Latte

* fixed test latte pipeline bugs

* fixed test latte pipeline bugs

* delete sh

* add doc for latte

* add licensing

* Move Transformer3DModelOutput to modeling_outputs

* give a default value to sample_size

* remove the einops dependency

* change norm2 for latte

* modify pipeline of latte

* update test for Latte

* modify some codes for latte

* modify for Latte pipeline

* modify for Latte pipeline

* modify for Latte pipeline

* modify for Latte pipeline

* modify for Latte pipeline

* modify for Latte pipeline

* modify for Latte pipeline

* modify for Latte pipeline

* modify for Latte pipeline

* modify for Latte pipeline

* modify for Latte pipeline

* modify for Latte pipeline

* modify for Latte pipeline

* modify for Latte pipeline

* modify for Latte pipeline

* modify for Latte pipeline

* modify for Latte pipeline

* modify for Latte pipeline

* modify for Latte pipeline

* modify for Latte pipeline

* modify for Latte pipeline

* modify for Latte pipeline

* modify for Latte pipeline

* modify for Latte pipeline

* modify for Latte pipeline

* modify for Latte pipeline

* modify for Latte pipeline

* video_length -> num_frames; update prepare_latents copied from

* make fix-copies

* make style

* typo: videe -> video

* update

* modify for Latte pipeline

* modify latte pipeline

* modify latte pipeline

* modify latte pipeline

* modify latte pipeline

* modify for Latte pipeline

* Delete .vscode directory

* make style

* make fix-copies

* add latte transformer 3d to docs _toctree.yml

* update example

* reduce frames for test

* fixed bug of _text_preprocessing

* set num frame to 1 for testing

* remove unuse print

* add text = self._clean_caption(text) again

---------
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>
Co-authored-by: YiYi Xu <yixu310@gmail.com>
Co-authored-by: Aryan <contact.aryanvs@gmail.com>
Co-authored-by: Aryan <aryan@huggingface.co>

.gitignore

@@ -175,4 +175,4 @@ tags
 .ruff_cache
 
 # wandb
-wandb
+wandb
\ No newline at end of file

docs/source/en/_toctree.yml

@@ -249,6 +249,8 @@
       title: DiTTransformer2DModel
     - local: api/models/hunyuan_transformer2d
       title: HunyuanDiT2DModel
+    - local: api/models/latte_transformer3d
+      title: LatteTransformer3DModel
     - local: api/models/lumina_nextdit2d
       title: LuminaNextDiT2DModel
     - local: api/models/transformer_temporal

docs/source/en/api/models/latte_transformer3d.md

+<!--Copyright 2024 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.
+-->
+
+## LatteTransformer3DModel
+
+A Diffusion Transformer model for 3D data from [Latte](https://github.com/Vchitect/Latte).
+
+## LatteTransformer3DModel
+
+[[autodoc]] LatteTransformer3DModel

src/diffusers/__init__.py

@@ -88,6 +88,7 @@ else:
             "HunyuanDiT2DMultiControlNetModel",
             "I2VGenXLUNet",
             "Kandinsky3UNet",
+            "LatteTransformer3DModel",
             "LuminaNextDiT2DModel",
             "ModelMixin",
             "MotionAdapter",
@@ -269,6 +270,7 @@ else:
             "KandinskyV22PriorPipeline",
             "LatentConsistencyModelImg2ImgPipeline",
             "LatentConsistencyModelPipeline",
+            "LattePipeline",
             "LDMTextToImagePipeline",
             "LEditsPPPipelineStableDiffusion",
             "LEditsPPPipelineStableDiffusionXL",
@@ -513,6 +515,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             HunyuanDiT2DMultiControlNetModel,
             I2VGenXLUNet,
             Kandinsky3UNet,
+            LatteTransformer3DModel,
             LuminaNextDiT2DModel,
             ModelMixin,
             MotionAdapter,
@@ -672,6 +675,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             KandinskyV22PriorPipeline,
             LatentConsistencyModelImg2ImgPipeline,
             LatentConsistencyModelPipeline,
+            LattePipeline,
             LDMTextToImagePipeline,
             LEditsPPPipelineStableDiffusion,
             LEditsPPPipelineStableDiffusionXL,

src/diffusers/models/__init__.py

@@ -41,6 +41,7 @@ if is_torch_available():
     _import_structure["transformers.dit_transformer_2d"] = ["DiTTransformer2DModel"]
     _import_structure["transformers.dual_transformer_2d"] = ["DualTransformer2DModel"]
     _import_structure["transformers.hunyuan_transformer_2d"] = ["HunyuanDiT2DModel"]
+    _import_structure["transformers.latte_transformer_3d"] = ["LatteTransformer3DModel"]
     _import_structure["transformers.lumina_nextdit2d"] = ["LuminaNextDiT2DModel"]
     _import_structure["transformers.pixart_transformer_2d"] = ["PixArtTransformer2DModel"]
     _import_structure["transformers.prior_transformer"] = ["PriorTransformer"]
@@ -86,6 +87,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             DiTTransformer2DModel,
             DualTransformer2DModel,
             HunyuanDiT2DModel,
+            LatteTransformer3DModel,
             LuminaNextDiT2DModel,
             PixArtTransformer2DModel,
             PriorTransformer,

src/diffusers/models/attention.py

@@ -359,7 +359,10 @@ class BasicTransformerBlock(nn.Module):
                 out_bias=attention_out_bias,
             )  # is self-attn if encoder_hidden_states is none
         else:
-            self.norm2 = None
+            if norm_type == "ada_norm_single":  # For Latte
+                self.norm2 = nn.LayerNorm(dim, norm_eps, norm_elementwise_affine)
+            else:
+                self.norm2 = None
             self.attn2 = None
 
         # 3. Feed-forward
@@ -439,7 +442,6 @@ class BasicTransformerBlock(nn.Module):
             ).chunk(6, dim=1)
             norm_hidden_states = self.norm1(hidden_states)
             norm_hidden_states = norm_hidden_states * (1 + scale_msa) + shift_msa
-            norm_hidden_states = norm_hidden_states.squeeze(1)
         else:
             raise ValueError("Incorrect norm used")
 
@@ -456,6 +458,7 @@ class BasicTransformerBlock(nn.Module):
             attention_mask=attention_mask,
             **cross_attention_kwargs,
         )
+
         if self.norm_type == "ada_norm_zero":
             attn_output = gate_msa.unsqueeze(1) * attn_output
         elif self.norm_type == "ada_norm_single":

src/diffusers/models/transformers/__init__.py

@@ -5,6 +5,7 @@ if is_torch_available():
     from .dit_transformer_2d import DiTTransformer2DModel
     from .dual_transformer_2d import DualTransformer2DModel
     from .hunyuan_transformer_2d import HunyuanDiT2DModel
+    from .latte_transformer_3d import LatteTransformer3DModel
     from .lumina_nextdit2d import LuminaNextDiT2DModel
     from .pixart_transformer_2d import PixArtTransformer2DModel
     from .prior_transformer import PriorTransformer

src/diffusers/models/transformers/latte_transformer_3d.py

+# Copyright 2024 the Latte Team 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 Optional
+
+import torch
+from torch import nn
+
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...models.embeddings import PixArtAlphaTextProjection, get_1d_sincos_pos_embed_from_grid
+from ..attention import BasicTransformerBlock
+from ..embeddings import PatchEmbed
+from ..modeling_outputs import Transformer2DModelOutput
+from ..modeling_utils import ModelMixin
+from ..normalization import AdaLayerNormSingle
+
+
+class LatteTransformer3DModel(ModelMixin, ConfigMixin):
+    _supports_gradient_checkpointing = True
+
+    """
+    A 3D Transformer model for video-like data, paper: https://arxiv.org/abs/2401.03048, offical code:
+    https://github.com/Vchitect/Latte
+
+    Parameters:
+        num_attention_heads (`int`, *optional*, defaults to 16): The number of heads to use for multi-head attention.
+        attention_head_dim (`int`, *optional*, defaults to 88): The number of channels in each head.
+        in_channels (`int`, *optional*):
+            The number of channels in the input.
+        out_channels (`int`, *optional*):
+            The number of channels in the output.
+        num_layers (`int`, *optional*, defaults to 1): The number of layers of Transformer blocks to use.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+        cross_attention_dim (`int`, *optional*): The number of `encoder_hidden_states` dimensions to use.
+        attention_bias (`bool`, *optional*):
+            Configure if the `TransformerBlocks` attention should contain a bias parameter.
+        sample_size (`int`, *optional*): The width of the latent images (specify if the input is **discrete**).
+            This is fixed during training since it is used to learn a number of position embeddings.
+        patch_size (`int`, *optional*):
+            The size of the patches to use in the patch embedding layer.
+        activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to use in feed-forward.
+        num_embeds_ada_norm ( `int`, *optional*):
+            The number of diffusion steps used during training. Pass if at least one of the norm_layers is
+            `AdaLayerNorm`. This is fixed during training since it is used to learn a number of embeddings that are
+            added to the hidden states. During inference, you can denoise for up to but not more steps than
+            `num_embeds_ada_norm`.
+        norm_type (`str`, *optional*, defaults to `"layer_norm"`):
+            The type of normalization to use. Options are `"layer_norm"` or `"ada_layer_norm"`.
+        norm_elementwise_affine (`bool`, *optional*, defaults to `True`):
+            Whether or not to use elementwise affine in normalization layers.
+        norm_eps (`float`, *optional*, defaults to 1e-5): The epsilon value to use in normalization layers.
+        caption_channels (`int`, *optional*):
+            The number of channels in the caption embeddings.
+        video_length (`int`, *optional*):
+            The number of frames in the video-like data.
+    """
+
+    @register_to_config
+    def __init__(
+        self,
+        num_attention_heads: int = 16,
+        attention_head_dim: int = 88,
+        in_channels: Optional[int] = None,
+        out_channels: Optional[int] = None,
+        num_layers: int = 1,
+        dropout: float = 0.0,
+        cross_attention_dim: Optional[int] = None,
+        attention_bias: bool = False,
+        sample_size: int = 64,
+        patch_size: Optional[int] = None,
+        activation_fn: str = "geglu",
+        num_embeds_ada_norm: Optional[int] = None,
+        norm_type: str = "layer_norm",
+        norm_elementwise_affine: bool = True,
+        norm_eps: float = 1e-5,
+        caption_channels: int = None,
+        video_length: int = 16,
+    ):
+        super().__init__()
+        inner_dim = num_attention_heads * attention_head_dim
+
+        # 1. Define input layers
+        self.height = sample_size
+        self.width = sample_size
+
+        interpolation_scale = self.config.sample_size // 64
+        interpolation_scale = max(interpolation_scale, 1)
+        self.pos_embed = PatchEmbed(
+            height=sample_size,
+            width=sample_size,
+            patch_size=patch_size,
+            in_channels=in_channels,
+            embed_dim=inner_dim,
+            interpolation_scale=interpolation_scale,
+        )
+
+        # 2. Define spatial transformers blocks
+        self.transformer_blocks = nn.ModuleList(
+            [
+                BasicTransformerBlock(
+                    inner_dim,
+                    num_attention_heads,
+                    attention_head_dim,
+                    dropout=dropout,
+                    cross_attention_dim=cross_attention_dim,
+                    activation_fn=activation_fn,
+                    num_embeds_ada_norm=num_embeds_ada_norm,
+                    attention_bias=attention_bias,
+                    norm_type=norm_type,
+                    norm_elementwise_affine=norm_elementwise_affine,
+                    norm_eps=norm_eps,
+                )
+                for d in range(num_layers)
+            ]
+        )
+
+        # 3. Define temporal transformers blocks
+        self.temporal_transformer_blocks = nn.ModuleList(
+            [
+                BasicTransformerBlock(
+                    inner_dim,
+                    num_attention_heads,
+                    attention_head_dim,
+                    dropout=dropout,
+                    cross_attention_dim=None,
+                    activation_fn=activation_fn,
+                    num_embeds_ada_norm=num_embeds_ada_norm,
+                    attention_bias=attention_bias,
+                    norm_type=norm_type,
+                    norm_elementwise_affine=norm_elementwise_affine,
+                    norm_eps=norm_eps,
+                )
+                for d in range(num_layers)
+            ]
+        )
+
+        # 4. Define output layers
+        self.out_channels = in_channels if out_channels is None else out_channels
+        self.norm_out = nn.LayerNorm(inner_dim, elementwise_affine=False, eps=1e-6)
+        self.scale_shift_table = nn.Parameter(torch.randn(2, inner_dim) / inner_dim**0.5)
+        self.proj_out = nn.Linear(inner_dim, patch_size * patch_size * self.out_channels)
+
+        # 5. Latte other blocks.
+        self.adaln_single = AdaLayerNormSingle(inner_dim, use_additional_conditions=False)
+        self.caption_projection = PixArtAlphaTextProjection(in_features=caption_channels, hidden_size=inner_dim)
+
+        # define temporal positional embedding
+        temp_pos_embed = get_1d_sincos_pos_embed_from_grid(
+            inner_dim, torch.arange(0, video_length).unsqueeze(1)
+        )  # 1152 hidden size
+        self.register_buffer("temp_pos_embed", torch.from_numpy(temp_pos_embed).float().unsqueeze(0), persistent=False)
+
+        self.gradient_checkpointing = False
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        self.gradient_checkpointing = value
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        timestep: Optional[torch.LongTensor] = None,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        enable_temporal_attentions: bool = True,
+        return_dict: bool = True,
+    ):
+        """
+        The [`LatteTransformer3DModel`] forward method.
+
+        Args:
+            hidden_states shape `(batch size, channel, num_frame, height, width)`:
+                Input `hidden_states`.
+            timestep ( `torch.LongTensor`, *optional*):
+                Used to indicate denoising step. Optional timestep to be applied as an embedding in `AdaLayerNorm`.
+            encoder_hidden_states ( `torch.FloatTensor` of shape `(batch size, sequence len, embed dims)`, *optional*):
+                Conditional embeddings for cross attention layer. If not given, cross-attention defaults to
+                self-attention.
+            encoder_attention_mask ( `torch.Tensor`, *optional*):
+                Cross-attention mask applied to `encoder_hidden_states`. Two formats supported:
+
+                    * Mask `(batcheight, sequence_length)` True = keep, False = discard.
+                    * Bias `(batcheight, 1, sequence_length)` 0 = keep, -10000 = discard.
+
+                If `ndim == 2`: will be interpreted as a mask, then converted into a bias consistent with the format
+                above. This bias will be added to the cross-attention scores.
+            enable_temporal_attentions:
+                (`bool`, *optional*, defaults to `True`): Whether to enable temporal attentions.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~models.unet_2d_condition.UNet2DConditionOutput`] instead of a plain
+                tuple.
+
+        Returns:
+            If `return_dict` is True, an [`~models.transformer_2d.Transformer2DModelOutput`] is returned, otherwise a
+            `tuple` where the first element is the sample tensor.
+        """
+
+        # Reshape hidden states
+        batch_size, channels, num_frame, height, width = hidden_states.shape
+        # batch_size channels num_frame height width -> (batch_size * num_frame) channels height width
+        hidden_states = hidden_states.permute(0, 2, 1, 3, 4).reshape(-1, channels, height, width)
+
+        # Input
+        height, width = (
+            hidden_states.shape[-2] // self.config.patch_size,
+            hidden_states.shape[-1] // self.config.patch_size,
+        )
+        num_patches = height * width
+
+        hidden_states = self.pos_embed(hidden_states)  # alrady add positional embeddings
+
+        added_cond_kwargs = {"resolution": None, "aspect_ratio": None}
+        timestep, embedded_timestep = self.adaln_single(
+            timestep, added_cond_kwargs=added_cond_kwargs, batch_size=batch_size, hidden_dtype=hidden_states.dtype
+        )
+
+        # Prepare text embeddings for spatial block
+        # batch_size num_tokens hidden_size -> (batch_size * num_frame) num_tokens hidden_size
+        encoder_hidden_states = self.caption_projection(encoder_hidden_states)  # 3 120 1152
+        encoder_hidden_states_spatial = encoder_hidden_states.repeat_interleave(num_frame, dim=0).view(
+            -1, encoder_hidden_states.shape[-2], encoder_hidden_states.shape[-1]
+        )
+
+        # Prepare timesteps for spatial and temporal block
+        timestep_spatial = timestep.repeat_interleave(num_frame, dim=0).view(-1, timestep.shape[-1])
+        timestep_temp = timestep.repeat_interleave(num_patches, dim=0).view(-1, timestep.shape[-1])
+
+        # Spatial and temporal transformer blocks
+        for i, (spatial_block, temp_block) in enumerate(
+            zip(self.transformer_blocks, self.temporal_transformer_blocks)
+        ):
+            if self.training and self.gradient_checkpointing:
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    spatial_block,
+                    hidden_states,
+                    None,  # attention_mask
+                    encoder_hidden_states_spatial,
+                    encoder_attention_mask,
+                    timestep_spatial,
+                    None,  # cross_attention_kwargs
+                    None,  # class_labels
+                    use_reentrant=False,
+                )
+            else:
+                hidden_states = spatial_block(
+                    hidden_states,
+                    None,  # attention_mask
+                    encoder_hidden_states_spatial,
+                    encoder_attention_mask,
+                    timestep_spatial,
+                    None,  # cross_attention_kwargs
+                    None,  # class_labels
+                )
+
+            if enable_temporal_attentions:
+                # (batch_size * num_frame) num_tokens hidden_size -> (batch_size * num_tokens) num_frame hidden_size
+                hidden_states = hidden_states.reshape(
+                    batch_size, -1, hidden_states.shape[-2], hidden_states.shape[-1]
+                ).permute(0, 2, 1, 3)
+                hidden_states = hidden_states.reshape(-1, hidden_states.shape[-2], hidden_states.shape[-1])
+
+                if i == 0 and num_frame > 1:
+                    hidden_states = hidden_states + self.temp_pos_embed
+
+                if self.training and self.gradient_checkpointing:
+                    hidden_states = torch.utils.checkpoint.checkpoint(
+                        temp_block,
+                        hidden_states,
+                        None,  # attention_mask
+                        None,  # encoder_hidden_states
+                        None,  # encoder_attention_mask
+                        timestep_temp,
+                        None,  # cross_attention_kwargs
+                        None,  # class_labels
+                        use_reentrant=False,
+                    )
+                else:
+                    hidden_states = temp_block(
+                        hidden_states,
+                        None,  # attention_mask
+                        None,  # encoder_hidden_states
+                        None,  # encoder_attention_mask
+                        timestep_temp,
+                        None,  # cross_attention_kwargs
+                        None,  # class_labels
+                    )
+
+                # (batch_size * num_tokens) num_frame hidden_size -> (batch_size * num_frame) num_tokens hidden_size
+                hidden_states = hidden_states.reshape(
+                    batch_size, -1, hidden_states.shape[-2], hidden_states.shape[-1]
+                ).permute(0, 2, 1, 3)
+                hidden_states = hidden_states.reshape(-1, hidden_states.shape[-2], hidden_states.shape[-1])
+
+        embedded_timestep = embedded_timestep.repeat_interleave(num_frame, dim=0).view(-1, embedded_timestep.shape[-1])
+        shift, scale = (self.scale_shift_table[None] + embedded_timestep[:, None]).chunk(2, dim=1)
+        hidden_states = self.norm_out(hidden_states)
+        # Modulation
+        hidden_states = hidden_states * (1 + scale) + shift
+        hidden_states = self.proj_out(hidden_states)
+
+        # unpatchify
+        if self.adaln_single is None:
+            height = width = int(hidden_states.shape[1] ** 0.5)
+        hidden_states = hidden_states.reshape(
+            shape=(-1, height, width, self.config.patch_size, self.config.patch_size, self.out_channels)
+        )
+        hidden_states = torch.einsum("nhwpqc->nchpwq", hidden_states)
+        output = hidden_states.reshape(
+            shape=(-1, self.out_channels, height * self.config.patch_size, width * self.config.patch_size)
+        )
+        output = output.reshape(batch_size, -1, output.shape[-3], output.shape[-2], output.shape[-1]).permute(
+            0, 2, 1, 3, 4
+        )
+
+        if not return_dict:
+            return (output,)
+
+        return Transformer2DModelOutput(sample=output)

src/diffusers/pipelines/__init__.py

@@ -209,6 +209,7 @@ else:
             "LEditsPPPipelineStableDiffusionXL",
         ]
     )
+    _import_structure["latte"] = ["LattePipeline"]
     _import_structure["lumina"] = ["LuminaText2ImgPipeline"]
     _import_structure["marigold"].extend(
         [
@@ -485,6 +486,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             LatentConsistencyModelPipeline,
         )
         from .latent_diffusion import LDMTextToImagePipeline
+        from .latte import LattePipeline
         from .ledits_pp import (
             LEditsPPDiffusionPipelineOutput,
             LEditsPPInversionPipelineOutput,

src/diffusers/pipelines/latte/__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_latte"] = ["LattePipeline"]
+
+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 *
+    else:
+        from .pipeline_latte import LattePipeline
+
+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/utils/dummy_pt_objects.py

@@ -197,6 +197,21 @@ class Kandinsky3UNet(metaclass=DummyObject):
         requires_backends(cls, ["torch"])
 
 
+class LatteTransformer3DModel(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 LuminaNextDiT2DModel(metaclass=DummyObject):
     _backends = ["torch"]
 

src/diffusers/utils/dummy_torch_and_transformers_objects.py

@@ -677,6 +677,21 @@ class LatentConsistencyModelPipeline(metaclass=DummyObject):
         requires_backends(cls, ["torch", "transformers"])
 
 
+class LattePipeline(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 LDMTextToImagePipeline(metaclass=DummyObject):
     _backends = ["torch", "transformers"]
 

tests/pipelines/latte/test_latte.py

+# coding=utf-8
+# Copyright 2024 Latte Team and 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 inspect
+import tempfile
+import unittest
+
+import numpy as np
+import torch
+from transformers import AutoTokenizer, T5EncoderModel
+
+from diffusers import (
+    AutoencoderKL,
+    DDIMScheduler,
+    LattePipeline,
+    LatteTransformer3DModel,
+)
+from diffusers.utils.testing_utils import (
+    enable_full_determinism,
+    numpy_cosine_similarity_distance,
+    require_torch_gpu,
+    slow,
+    torch_device,
+)
+
+from ..pipeline_params import TEXT_TO_IMAGE_BATCH_PARAMS, TEXT_TO_IMAGE_IMAGE_PARAMS, TEXT_TO_IMAGE_PARAMS
+from ..test_pipelines_common import PipelineTesterMixin, to_np
+
+
+enable_full_determinism()
+
+
+class LattePipelineFastTests(PipelineTesterMixin, unittest.TestCase):
+    pipeline_class = LattePipeline
+    params = TEXT_TO_IMAGE_PARAMS - {"cross_attention_kwargs"}
+    batch_params = TEXT_TO_IMAGE_BATCH_PARAMS
+    image_params = TEXT_TO_IMAGE_IMAGE_PARAMS
+    image_latents_params = TEXT_TO_IMAGE_IMAGE_PARAMS
+
+    required_optional_params = PipelineTesterMixin.required_optional_params
+
+    def get_dummy_components(self):
+        torch.manual_seed(0)
+        transformer = LatteTransformer3DModel(
+            sample_size=8,
+            num_layers=1,
+            patch_size=2,
+            attention_head_dim=8,
+            num_attention_heads=3,
+            caption_channels=32,
+            in_channels=4,
+            cross_attention_dim=24,
+            out_channels=8,
+            attention_bias=True,
+            activation_fn="gelu-approximate",
+            num_embeds_ada_norm=1000,
+            norm_type="ada_norm_single",
+            norm_elementwise_affine=False,
+            norm_eps=1e-6,
+        )
+        torch.manual_seed(0)
+        vae = AutoencoderKL()
+
+        scheduler = DDIMScheduler()
+        text_encoder = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+
+        tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")
+
+        components = {
+            "transformer": transformer.eval(),
+            "vae": vae.eval(),
+            "scheduler": scheduler,
+            "text_encoder": text_encoder.eval(),
+            "tokenizer": tokenizer,
+        }
+        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)
+        inputs = {
+            "prompt": "A painting of a squirrel eating a burger",
+            "negative_prompt": "low quality",
+            "generator": generator,
+            "num_inference_steps": 2,
+            "guidance_scale": 5.0,
+            "height": 8,
+            "width": 8,
+            "video_length": 1,
+            "output_type": "pt",
+            "clean_caption": False,
+        }
+        return inputs
+
+    def test_inference(self):
+        device = "cpu"
+
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+        pipe.to(device)
+        pipe.set_progress_bar_config(disable=None)
+
+        inputs = self.get_dummy_inputs(device)
+        video = pipe(**inputs).frames
+        generated_video = video[0]
+
+        self.assertEqual(generated_video.shape, (1, 3, 8, 8))
+        expected_video = torch.randn(1, 3, 8, 8)
+        max_diff = np.abs(generated_video - expected_video).max()
+        self.assertLessEqual(max_diff, 1e10)
+
+    def test_callback_inputs(self):
+        sig = inspect.signature(self.pipeline_class.__call__)
+        has_callback_tensor_inputs = "callback_on_step_end_tensor_inputs" in sig.parameters
+        has_callback_step_end = "callback_on_step_end" in sig.parameters
+
+        if not (has_callback_tensor_inputs and has_callback_step_end):
+            return
+
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+        pipe = pipe.to(torch_device)
+        pipe.set_progress_bar_config(disable=None)
+        self.assertTrue(
+            hasattr(pipe, "_callback_tensor_inputs"),
+            f" {self.pipeline_class} should have `_callback_tensor_inputs` that defines a list of tensor variables its callback function can use as inputs",
+        )
+
+        def callback_inputs_subset(pipe, i, t, callback_kwargs):
+            # iterate over callback args
+            for tensor_name, tensor_value in callback_kwargs.items():
+                # check that we're only passing in allowed tensor inputs
+                assert tensor_name in pipe._callback_tensor_inputs
+
+            return callback_kwargs
+
+        def callback_inputs_all(pipe, i, t, callback_kwargs):
+            for tensor_name in pipe._callback_tensor_inputs:
+                assert tensor_name in callback_kwargs
+
+            # iterate over callback args
+            for tensor_name, tensor_value in callback_kwargs.items():
+                # check that we're only passing in allowed tensor inputs
+                assert tensor_name in pipe._callback_tensor_inputs
+
+            return callback_kwargs
+
+        inputs = self.get_dummy_inputs(torch_device)
+
+        # Test passing in a subset
+        inputs["callback_on_step_end"] = callback_inputs_subset
+        inputs["callback_on_step_end_tensor_inputs"] = ["latents"]
+        output = pipe(**inputs)[0]
+
+        # Test passing in a everything
+        inputs["callback_on_step_end"] = callback_inputs_all
+        inputs["callback_on_step_end_tensor_inputs"] = pipe._callback_tensor_inputs
+        output = pipe(**inputs)[0]
+
+        def callback_inputs_change_tensor(pipe, i, t, callback_kwargs):
+            is_last = i == (pipe.num_timesteps - 1)
+            if is_last:
+                callback_kwargs["latents"] = torch.zeros_like(callback_kwargs["latents"])
+            return callback_kwargs
+
+        inputs["callback_on_step_end"] = callback_inputs_change_tensor
+        inputs["callback_on_step_end_tensor_inputs"] = pipe._callback_tensor_inputs
+        output = pipe(**inputs)[0]
+        assert output.abs().sum() < 1e10
+
+    def test_inference_batch_single_identical(self):
+        self._test_inference_batch_single_identical(batch_size=3, expected_max_diff=1e-3)
+
+    def test_attention_slicing_forward_pass(self):
+        pass
+
+    def test_save_load_optional_components(self):
+        if not hasattr(self.pipeline_class, "_optional_components"):
+            return
+
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+
+        for component in pipe.components.values():
+            if hasattr(component, "set_default_attn_processor"):
+                component.set_default_attn_processor()
+        pipe.to(torch_device)
+        pipe.set_progress_bar_config(disable=None)
+
+        inputs = self.get_dummy_inputs(torch_device)
+
+        prompt = inputs["prompt"]
+        generator = inputs["generator"]
+
+        (
+            prompt_embeds,
+            negative_prompt_embeds,
+        ) = pipe.encode_prompt(prompt)
+
+        # inputs with prompt converted to embeddings
+        inputs = {
+            "prompt_embeds": prompt_embeds,
+            "negative_prompt": None,
+            "negative_prompt_embeds": negative_prompt_embeds,
+            "generator": generator,
+            "num_inference_steps": 2,
+            "guidance_scale": 5.0,
+            "height": 8,
+            "width": 8,
+            "video_length": 1,
+            "mask_feature": False,
+            "output_type": "pt",
+            "clean_caption": False,
+        }
+
+        # set all optional components to None
+        for optional_component in pipe._optional_components:
+            setattr(pipe, optional_component, None)
+
+        output = pipe(**inputs)[0]
+
+        with tempfile.TemporaryDirectory() as tmpdir:
+            pipe.save_pretrained(tmpdir, safe_serialization=False)
+            pipe_loaded = self.pipeline_class.from_pretrained(tmpdir)
+            pipe_loaded.to(torch_device)
+
+            for component in pipe_loaded.components.values():
+                if hasattr(component, "set_default_attn_processor"):
+                    component.set_default_attn_processor()
+
+            pipe_loaded.set_progress_bar_config(disable=None)
+
+        for optional_component in pipe._optional_components:
+            self.assertTrue(
+                getattr(pipe_loaded, optional_component) is None,
+                f"`{optional_component}` did not stay set to None after loading.",
+            )
+
+        output_loaded = pipe_loaded(**inputs)[0]
+
+        max_diff = np.abs(to_np(output) - to_np(output_loaded)).max()
+        self.assertLess(max_diff, 1.0)
+
+
+@slow
+@require_torch_gpu
+class LattePipelineIntegrationTests(unittest.TestCase):
+    prompt = "A painting of a squirrel eating a burger."
+
+    def setUp(self):
+        super().setUp()
+        gc.collect()
+        torch.cuda.empty_cache()
+
+    def tearDown(self):
+        super().tearDown()
+        gc.collect()
+        torch.cuda.empty_cache()
+
+    def test_latte(self):
+        generator = torch.Generator("cpu").manual_seed(0)
+
+        pipe = LattePipeline.from_pretrained("maxin-cn/Latte-1", torch_dtype=torch.float16)
+        pipe.enable_model_cpu_offload()
+        prompt = self.prompt
+
+        videos = pipe(
+            prompt=prompt,
+            height=512,
+            width=512,
+            generator=generator,
+            num_inference_steps=2,
+            clean_caption=False,
+        ).frames
+
+        video = videos[0]
+        expected_video = torch.randn(1, 512, 512, 3).numpy()
+
+        max_diff = numpy_cosine_similarity_distance(video.flatten(), expected_video)
+        assert max_diff < 1e-3, f"Max diff is too high. got {video.flatten()}"