Merge branch 'main' of https://github.com/huggingface/diffusers into main

docs/source/api/pipelines/stable_diffusion.mdx

@@ -95,3 +95,10 @@ If you want to use all possible use cases in a single `DiffusionPipeline` you ca
 	- __call__
 	- enable_attention_slicing
 	- disable_attention_slicing
+
+
+## StableDiffusionUpscalePipeline
+[[autodoc]] StableDiffusionUpscalePipeline
+	- __call__
+	- enable_attention_slicing
+	- disable_attention_slicing

src/diffusers/__init__.py

@@ -75,6 +75,7 @@ if is_torch_available() and is_transformers_available():
         StableDiffusionInpaintPipelineLegacy,
         StableDiffusionPipeline,
         StableDiffusionPipelineSafe,
+        StableDiffusionUpscalePipeline,
         VersatileDiffusionDualGuidedPipeline,
         VersatileDiffusionImageVariationPipeline,
         VersatileDiffusionPipeline,

src/diffusers/pipeline_utils.py

@@ -554,7 +554,9 @@ class DiffusionPipeline(ConfigMixin):
         init_dict = {k: v for k, v in init_dict.items() if load_module(k, v)}
 
         if len(unused_kwargs) > 0:
-            logger.warning(f"Keyword arguments {unused_kwargs} not recognized.")
+            logger.warning(
+                f"Keyword arguments {unused_kwargs} are not expected by {pipeline_class.__name__} and will be ignored."
+            )
 
         # import it here to avoid circular import
         from diffusers import pipelines
@@ -680,8 +682,8 @@ class DiffusionPipeline(ConfigMixin):
     @staticmethod
     def _get_signature_keys(obj):
         parameters = inspect.signature(obj.__init__).parameters
-        required_parameters = {k: v for k, v in parameters.items() if v.default is not True}
-        optional_parameters = set({k for k, v in parameters.items() if v.default is True})
+        required_parameters = {k: v for k, v in parameters.items() if v.default == inspect._empty}
+        optional_parameters = set({k for k, v in parameters.items() if v.default != inspect._empty})
         expected_modules = set(required_parameters.keys()) - set(["self"])
         return expected_modules, optional_parameters
 

src/diffusers/pipelines/__init__.py

@@ -24,6 +24,7 @@ if is_torch_available() and is_transformers_available():
         StableDiffusionInpaintPipeline,
         StableDiffusionInpaintPipelineLegacy,
         StableDiffusionPipeline,
+        StableDiffusionUpscalePipeline,
     )
     from .stable_diffusion_safe import StableDiffusionPipelineSafe
     from .versatile_diffusion import (

src/diffusers/pipelines/stable_diffusion/__init__.py

@@ -40,6 +40,7 @@ if is_transformers_available() and is_torch_available():
     from .pipeline_stable_diffusion_img2img import StableDiffusionImg2ImgPipeline
     from .pipeline_stable_diffusion_inpaint import StableDiffusionInpaintPipeline
     from .pipeline_stable_diffusion_inpaint_legacy import StableDiffusionInpaintPipelineLegacy
+    from .pipeline_stable_diffusion_upscale import StableDiffusionUpscalePipeline
     from .safety_checker import StableDiffusionSafetyChecker
 
 if is_transformers_available() and is_torch_available() and is_transformers_version(">=", "4.25.0.dev0"):

src/diffusers/utils/dummy_torch_and_transformers_objects.py

@@ -154,6 +154,21 @@ class StableDiffusionPipelineSafe(metaclass=DummyObject):
         requires_backends(cls, ["torch", "transformers"])
 
 
+class StableDiffusionUpscalePipeline(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 VersatileDiffusionDualGuidedPipeline(metaclass=DummyObject):
     _backends = ["torch", "transformers"]
 

tests/pipelines/stable_diffusion_2/test_stable_diffusion.py

@@ -34,7 +34,7 @@ from diffusers import (
 )
 from diffusers.utils import load_numpy, slow, torch_device
 from diffusers.utils.testing_utils import CaptureLogger, require_torch_gpu
-from transformers import CLIPFeatureExtractor, CLIPTextConfig, CLIPTextModel, CLIPTokenizer
+from transformers import CLIPTextConfig, CLIPTextModel, CLIPTokenizer
 
 from ...test_pipelines_common import PipelineTesterMixin
 
@@ -100,21 +100,6 @@ class StableDiffusion2PipelineFastTests(PipelineTesterMixin, unittest.TestCase):
         )
         return CLIPTextModel(config)
 
-    @property
-    def dummy_extractor(self):
-        def extract(*args, **kwargs):
-            class Out:
-                def __init__(self):
-                    self.pixel_values = torch.ones([0])
-
-                def to(self, device):
-                    self.pixel_values.to(device)
-                    return self
-
-            return Out()
-
-        return extract
-
     def test_save_pretrained_from_pretrained(self):
         device = "cpu"  # ensure determinism for the device-dependent torch.Generator
         unet = self.dummy_cond_unet
@@ -129,7 +114,6 @@ class StableDiffusion2PipelineFastTests(PipelineTesterMixin, unittest.TestCase):
         vae = self.dummy_vae
         bert = self.dummy_text_encoder
         tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
-        feature_extractor = CLIPFeatureExtractor.from_pretrained("hf-internal-testing/tiny-random-clip")
 
         # make sure here that pndm scheduler skips prk
         sd_pipe = StableDiffusionPipeline(
@@ -139,7 +123,8 @@ class StableDiffusion2PipelineFastTests(PipelineTesterMixin, unittest.TestCase):
             text_encoder=bert,
             tokenizer=tokenizer,
             safety_checker=None,
-            feature_extractor=feature_extractor,
+            feature_extractor=None,
+            requires_safety_checker=False,
         )
         sd_pipe = sd_pipe.to(device)
         sd_pipe.set_progress_bar_config(disable=None)
@@ -185,7 +170,8 @@ class StableDiffusion2PipelineFastTests(PipelineTesterMixin, unittest.TestCase):
             text_encoder=bert,
             tokenizer=tokenizer,
             safety_checker=None,
-            feature_extractor=self.dummy_extractor,
+            feature_extractor=None,
+            requires_safety_checker=False,
         )
         sd_pipe = sd_pipe.to(device)
         sd_pipe.set_progress_bar_config(disable=None)
@@ -231,7 +217,8 @@ class StableDiffusion2PipelineFastTests(PipelineTesterMixin, unittest.TestCase):
             text_encoder=bert,
             tokenizer=tokenizer,
             safety_checker=None,
-            feature_extractor=self.dummy_extractor,
+            feature_extractor=None,
+            requires_safety_checker=False,
         )
         sd_pipe = sd_pipe.to(device)
         sd_pipe.set_progress_bar_config(disable=None)
@@ -276,7 +263,8 @@ class StableDiffusion2PipelineFastTests(PipelineTesterMixin, unittest.TestCase):
             text_encoder=bert,
             tokenizer=tokenizer,
             safety_checker=None,
-            feature_extractor=self.dummy_extractor,
+            feature_extractor=None,
+            requires_safety_checker=False,
         )
         sd_pipe = sd_pipe.to(device)
         sd_pipe.set_progress_bar_config(disable=None)
@@ -321,7 +309,8 @@ class StableDiffusion2PipelineFastTests(PipelineTesterMixin, unittest.TestCase):
             text_encoder=bert,
             tokenizer=tokenizer,
             safety_checker=None,
-            feature_extractor=self.dummy_extractor,
+            feature_extractor=None,
+            requires_safety_checker=False,
         )
         sd_pipe = sd_pipe.to(device)
         sd_pipe.set_progress_bar_config(disable=None)
@@ -366,7 +355,8 @@ class StableDiffusion2PipelineFastTests(PipelineTesterMixin, unittest.TestCase):
             text_encoder=bert,
             tokenizer=tokenizer,
             safety_checker=None,
-            feature_extractor=self.dummy_extractor,
+            feature_extractor=None,
+            requires_safety_checker=False,
         )
         sd_pipe = sd_pipe.to(device)
         sd_pipe.set_progress_bar_config(disable=None)
@@ -411,7 +401,8 @@ class StableDiffusion2PipelineFastTests(PipelineTesterMixin, unittest.TestCase):
             text_encoder=bert,
             tokenizer=tokenizer,
             safety_checker=None,
-            feature_extractor=self.dummy_extractor,
+            feature_extractor=None,
+            requires_safety_checker=False,
         )
         sd_pipe = sd_pipe.to(device)
         sd_pipe.set_progress_bar_config(disable=None)
@@ -449,7 +440,8 @@ class StableDiffusion2PipelineFastTests(PipelineTesterMixin, unittest.TestCase):
             text_encoder=bert,
             tokenizer=tokenizer,
             safety_checker=None,
-            feature_extractor=self.dummy_extractor,
+            feature_extractor=None,
+            requires_safety_checker=False,
         )
         sd_pipe = sd_pipe.to(torch_device)
         sd_pipe.set_progress_bar_config(disable=None)
@@ -475,7 +467,8 @@ class StableDiffusion2PipelineFastTests(PipelineTesterMixin, unittest.TestCase):
             text_encoder=bert,
             tokenizer=tokenizer,
             safety_checker=None,
-            feature_extractor=self.dummy_extractor,
+            feature_extractor=None,
+            requires_safety_checker=False,
         )
         sd_pipe = sd_pipe.to(torch_device)
         sd_pipe.set_progress_bar_config(disable=None)
@@ -572,7 +565,7 @@ class StableDiffusion2PipelineIntegrationTests(unittest.TestCase):
         expected_slice = np.array([0.0548, 0.0626, 0.0612, 0.0611, 0.0706, 0.0586, 0.0843, 0.0333, 0.1197])
         assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2
 
-    def test_stable_diffusion_memory_chunking(self):
+    def test_stable_diffusion_attention_slicing(self):
         torch.cuda.reset_peak_memory_stats()
         model_id = "stabilityai/stable-diffusion-2-base"
         pipe = StableDiffusionPipeline.from_pretrained(model_id, revision="fp16", torch_dtype=torch.float16)
@@ -651,7 +644,7 @@ class StableDiffusion2PipelineIntegrationTests(unittest.TestCase):
         prompt = "astronaut riding a horse"
 
         generator = torch.Generator(device=torch_device).manual_seed(0)
-        output = pipe(prompt=prompt, strength=0.75, guidance_scale=7.5, generator=generator, output_type="np")
+        output = pipe(prompt=prompt, guidance_scale=7.5, generator=generator, output_type="np")
         image = output.images[0]
 
         assert image.shape == (512, 512, 3)

tests/pipelines/stable_diffusion_2/test_stable_diffusion_upscale.py

+# coding=utf-8
+# Copyright 2022 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 diffusers import AutoencoderKL, DDIMScheduler, DDPMScheduler, StableDiffusionUpscalePipeline, UNet2DConditionModel
+from diffusers.utils import floats_tensor, load_image, load_numpy, slow, torch_device
+from diffusers.utils.testing_utils import require_torch_gpu
+from PIL import Image
+from transformers import CLIPTextConfig, CLIPTextModel, CLIPTokenizer
+
+from ...test_pipelines_common import PipelineTesterMixin
+
+
+torch.backends.cuda.matmul.allow_tf32 = False
+
+
+class StableDiffusionUpscalePipelineFastTests(PipelineTesterMixin, unittest.TestCase):
+    def tearDown(self):
+        # clean up the VRAM after each test
+        super().tearDown()
+        gc.collect()
+        torch.cuda.empty_cache()
+
+    @property
+    def dummy_image(self):
+        batch_size = 1
+        num_channels = 3
+        sizes = (32, 32)
+
+        image = floats_tensor((batch_size, num_channels) + sizes, rng=random.Random(0)).to(torch_device)
+        return image
+
+    @property
+    def dummy_cond_unet_upscale(self):
+        torch.manual_seed(0)
+        model = UNet2DConditionModel(
+            block_out_channels=(32, 32, 64),
+            layers_per_block=2,
+            sample_size=32,
+            in_channels=7,
+            out_channels=4,
+            down_block_types=("DownBlock2D", "CrossAttnDownBlock2D", "CrossAttnDownBlock2D"),
+            up_block_types=("CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "UpBlock2D"),
+            cross_attention_dim=32,
+            # SD2-specific config below
+            attention_head_dim=8,
+            use_linear_projection=True,
+            only_cross_attention=(True, True, False),
+            num_class_embeds=100,
+        )
+        return model
+
+    @property
+    def dummy_vae(self):
+        torch.manual_seed(0)
+        model = AutoencoderKL(
+            block_out_channels=[32, 32, 64],
+            in_channels=3,
+            out_channels=3,
+            down_block_types=["DownEncoderBlock2D", "DownEncoderBlock2D", "DownEncoderBlock2D"],
+            up_block_types=["UpDecoderBlock2D", "UpDecoderBlock2D", "UpDecoderBlock2D"],
+            latent_channels=4,
+        )
+        return model
+
+    @property
+    def dummy_text_encoder(self):
+        torch.manual_seed(0)
+        config = CLIPTextConfig(
+            bos_token_id=0,
+            eos_token_id=2,
+            hidden_size=32,
+            intermediate_size=37,
+            layer_norm_eps=1e-05,
+            num_attention_heads=4,
+            num_hidden_layers=5,
+            pad_token_id=1,
+            vocab_size=1000,
+            # SD2-specific config below
+            hidden_act="gelu",
+            projection_dim=512,
+        )
+        return CLIPTextModel(config)
+
+    def test_stable_diffusion_upscale(self):
+        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
+        unet = self.dummy_cond_unet_upscale
+        low_res_scheduler = DDPMScheduler()
+        scheduler = DDIMScheduler(prediction_type="v_prediction")
+        vae = self.dummy_vae
+        text_encoder = self.dummy_text_encoder
+        tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
+
+        image = self.dummy_image.cpu().permute(0, 2, 3, 1)[0]
+        low_res_image = Image.fromarray(np.uint8(image)).convert("RGB").resize((64, 64))
+
+        # make sure here that pndm scheduler skips prk
+        sd_pipe = StableDiffusionUpscalePipeline(
+            unet=unet,
+            low_res_scheduler=low_res_scheduler,
+            scheduler=scheduler,
+            vae=vae,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            max_noise_level=350,
+        )
+        sd_pipe = sd_pipe.to(device)
+        sd_pipe.set_progress_bar_config(disable=None)
+
+        prompt = "A painting of a squirrel eating a burger"
+        generator = torch.Generator(device=device).manual_seed(0)
+        output = sd_pipe(
+            [prompt],
+            image=low_res_image,
+            generator=generator,
+            guidance_scale=6.0,
+            noise_level=20,
+            num_inference_steps=2,
+            output_type="np",
+        )
+
+        image = output.images
+
+        generator = torch.Generator(device=device).manual_seed(0)
+        image_from_tuple = sd_pipe(
+            [prompt],
+            image=low_res_image,
+            generator=generator,
+            guidance_scale=6.0,
+            noise_level=20,
+            num_inference_steps=2,
+            output_type="np",
+            return_dict=False,
+        )[0]
+
+        image_slice = image[0, -3:, -3:, -1]
+        image_from_tuple_slice = image_from_tuple[0, -3:, -3:, -1]
+
+        expected_height_width = low_res_image.size[0] * 4
+        assert image.shape == (1, expected_height_width, expected_height_width, 3)
+        expected_slice = np.array([0.2562, 0.3606, 0.4204, 0.4469, 0.4822, 0.4647, 0.5315, 0.5748, 0.5606])
+
+        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2
+        assert np.abs(image_from_tuple_slice.flatten() - expected_slice).max() < 1e-2
+
+    @unittest.skipIf(torch_device != "cuda", "This test requires a GPU")
+    def test_stable_diffusion_upscale_fp16(self):
+        """Test that stable diffusion upscale works with fp16"""
+        unet = self.dummy_cond_unet_upscale
+        low_res_scheduler = DDPMScheduler()
+        scheduler = DDIMScheduler(prediction_type="v_prediction")
+        vae = self.dummy_vae
+        text_encoder = self.dummy_text_encoder
+        tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
+
+        image = self.dummy_image.cpu().permute(0, 2, 3, 1)[0]
+        low_res_image = Image.fromarray(np.uint8(image)).convert("RGB").resize((64, 64))
+
+        # put models in fp16, except vae as it overflows in fp16
+        unet = unet.half()
+        text_encoder = text_encoder.half()
+
+        # make sure here that pndm scheduler skips prk
+        sd_pipe = StableDiffusionUpscalePipeline(
+            unet=unet,
+            low_res_scheduler=low_res_scheduler,
+            scheduler=scheduler,
+            vae=vae,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            max_noise_level=350,
+        )
+        sd_pipe = sd_pipe.to(torch_device)
+        sd_pipe.set_progress_bar_config(disable=None)
+
+        prompt = "A painting of a squirrel eating a burger"
+        generator = torch.Generator(device=torch_device).manual_seed(0)
+        image = sd_pipe(
+            [prompt],
+            image=low_res_image,
+            generator=generator,
+            num_inference_steps=2,
+            output_type="np",
+        ).images
+
+        expected_height_width = low_res_image.size[0] * 4
+        assert image.shape == (1, expected_height_width, expected_height_width, 3)
+
+
+@slow
+@require_torch_gpu
+class StableDiffusionUpscalePipelineIntegrationTests(unittest.TestCase):
+    def tearDown(self):
+        # clean up the VRAM after each test
+        super().tearDown()
+        gc.collect()
+        torch.cuda.empty_cache()
+
+    def test_stable_diffusion_upscale_pipeline(self):
+        image = load_image(
+            "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main"
+            "/sd2-upscale/low_res_cat.png"
+        )
+        expected_image = load_numpy(
+            "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/sd2-upscale"
+            "/upsampled_cat.npy"
+        )
+
+        model_id = "stabilityai/stable-diffusion-x4-upscaler"
+        pipe = StableDiffusionUpscalePipeline.from_pretrained(model_id)
+        pipe.to(torch_device)
+        pipe.set_progress_bar_config(disable=None)
+        pipe.enable_attention_slicing()
+
+        prompt = "a cat sitting on a park bench"
+
+        generator = torch.Generator(device=torch_device).manual_seed(0)
+        output = pipe(
+            prompt=prompt,
+            image=image,
+            generator=generator,
+            output_type="np",
+        )
+        image = output.images[0]
+
+        assert image.shape == (512, 512, 3)
+        assert np.abs(expected_image - image).max() < 1e-3
+
+    def test_stable_diffusion_upscale_pipeline_fp16(self):
+        image = load_image(
+            "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main"
+            "/sd2-upscale/low_res_cat.png"
+        )
+        expected_image = load_numpy(
+            "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/sd2-upscale"
+            "/upsampled_cat_fp16.npy"
+        )
+
+        model_id = "stabilityai/stable-diffusion-x4-upscaler"
+        pipe = StableDiffusionUpscalePipeline.from_pretrained(
+            model_id,
+            revision="fp16",
+            torch_dtype=torch.float16,
+        )
+        pipe.to(torch_device)
+        pipe.set_progress_bar_config(disable=None)
+        pipe.enable_attention_slicing()
+
+        prompt = "a cat sitting on a park bench"
+
+        generator = torch.Generator(device=torch_device).manual_seed(0)
+        output = pipe(
+            prompt=prompt,
+            image=image,
+            generator=generator,
+            output_type="np",
+        )
+        image = output.images[0]
+
+        assert image.shape == (512, 512, 3)
+        assert np.abs(expected_image - image).max() < 5e-1
+
+    def test_stable_diffusion_pipeline_with_sequential_cpu_offloading(self):
+        torch.cuda.empty_cache()
+        torch.cuda.reset_max_memory_allocated()
+        torch.cuda.reset_peak_memory_stats()
+
+        image = load_image(
+            "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main"
+            "/sd2-upscale/low_res_cat.png"
+        )
+
+        model_id = "stabilityai/stable-diffusion-x4-upscaler"
+        pipe = StableDiffusionUpscalePipeline.from_pretrained(
+            model_id,
+            revision="fp16",
+            torch_dtype=torch.float16,
+        )
+        pipe.to(torch_device)
+        pipe.set_progress_bar_config(disable=None)
+        pipe.enable_attention_slicing(1)
+        pipe.enable_sequential_cpu_offload()
+
+        prompt = "a cat sitting on a park bench"
+
+        generator = torch.Generator(device=torch_device).manual_seed(0)
+        _ = pipe(
+            prompt=prompt,
+            image=image,
+            generator=generator,
+            num_inference_steps=5,
+            output_type="np",
+        )
+
+        mem_bytes = torch.cuda.max_memory_allocated()
+        # make sure that less than 2.65 GB is allocated
+        assert mem_bytes < 2.65 * 10**9

tests/pipelines/stable_diffusion_2/test_stable_diffusion_v_pred.py

+# coding=utf-8
+# Copyright 2022 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 time
+import unittest
+
+import numpy as np
+import torch
+
+from diffusers import (
+    AutoencoderKL,
+    DDIMScheduler,
+    DPMSolverMultistepScheduler,
+    EulerDiscreteScheduler,
+    StableDiffusionPipeline,
+    UNet2DConditionModel,
+)
+from diffusers.utils import load_numpy, slow, torch_device
+from diffusers.utils.testing_utils import require_torch_gpu
+from transformers import CLIPTextConfig, CLIPTextModel, CLIPTokenizer
+
+from ...test_pipelines_common import PipelineTesterMixin
+
+
+torch.backends.cuda.matmul.allow_tf32 = False
+
+
+class StableDiffusion2VPredictionPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
+    def tearDown(self):
+        # clean up the VRAM after each test
+        super().tearDown()
+        gc.collect()
+        torch.cuda.empty_cache()
+
+    @property
+    def dummy_cond_unet(self):
+        torch.manual_seed(0)
+        model = UNet2DConditionModel(
+            block_out_channels=(32, 64),
+            layers_per_block=2,
+            sample_size=32,
+            in_channels=4,
+            out_channels=4,
+            down_block_types=("DownBlock2D", "CrossAttnDownBlock2D"),
+            up_block_types=("CrossAttnUpBlock2D", "UpBlock2D"),
+            cross_attention_dim=32,
+            # SD2-specific config below
+            attention_head_dim=(2, 4, 8, 8),
+            use_linear_projection=True,
+        )
+        return model
+
+    @property
+    def dummy_vae(self):
+        torch.manual_seed(0)
+        model = AutoencoderKL(
+            block_out_channels=[32, 64],
+            in_channels=3,
+            out_channels=3,
+            down_block_types=["DownEncoderBlock2D", "DownEncoderBlock2D"],
+            up_block_types=["UpDecoderBlock2D", "UpDecoderBlock2D"],
+            latent_channels=4,
+            sample_size=128,
+        )
+        return model
+
+    @property
+    def dummy_text_encoder(self):
+        torch.manual_seed(0)
+        config = CLIPTextConfig(
+            bos_token_id=0,
+            eos_token_id=2,
+            hidden_size=32,
+            intermediate_size=37,
+            layer_norm_eps=1e-05,
+            num_attention_heads=4,
+            num_hidden_layers=5,
+            pad_token_id=1,
+            vocab_size=1000,
+            # SD2-specific config below
+            hidden_act="gelu",
+            projection_dim=64,
+        )
+        return CLIPTextModel(config)
+
+    def test_stable_diffusion_v_pred_ddim(self):
+        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
+        unet = self.dummy_cond_unet
+        scheduler = DDIMScheduler(
+            beta_start=0.00085,
+            beta_end=0.012,
+            beta_schedule="scaled_linear",
+            clip_sample=False,
+            set_alpha_to_one=False,
+            prediction_type="v_prediction",
+        )
+
+        vae = self.dummy_vae
+        bert = self.dummy_text_encoder
+        tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
+
+        # make sure here that pndm scheduler skips prk
+        sd_pipe = StableDiffusionPipeline(
+            unet=unet,
+            scheduler=scheduler,
+            vae=vae,
+            text_encoder=bert,
+            tokenizer=tokenizer,
+            safety_checker=None,
+            feature_extractor=None,
+            requires_safety_checker=False,
+        )
+        sd_pipe = sd_pipe.to(device)
+        sd_pipe.set_progress_bar_config(disable=None)
+
+        prompt = "A painting of a squirrel eating a burger"
+
+        generator = torch.Generator(device=device).manual_seed(0)
+        output = sd_pipe([prompt], generator=generator, guidance_scale=6.0, num_inference_steps=2, output_type="np")
+        image = output.images
+
+        generator = torch.Generator(device=device).manual_seed(0)
+        image_from_tuple = sd_pipe(
+            [prompt],
+            generator=generator,
+            guidance_scale=6.0,
+            num_inference_steps=2,
+            output_type="np",
+            return_dict=False,
+        )[0]
+
+        image_slice = image[0, -3:, -3:, -1]
+        image_from_tuple_slice = image_from_tuple[0, -3:, -3:, -1]
+
+        assert image.shape == (1, 64, 64, 3)
+        expected_slice = np.array([0.6424, 0.6109, 0.494, 0.5088, 0.4984, 0.4525, 0.5059, 0.5068, 0.4474])
+
+        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2
+        assert np.abs(image_from_tuple_slice.flatten() - expected_slice).max() < 1e-2
+
+    def test_stable_diffusion_v_pred_k_euler(self):
+        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
+        unet = self.dummy_cond_unet
+        scheduler = EulerDiscreteScheduler(
+            beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear", prediction_type="v_prediction"
+        )
+        vae = self.dummy_vae
+        bert = self.dummy_text_encoder
+        tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
+
+        # make sure here that pndm scheduler skips prk
+        sd_pipe = StableDiffusionPipeline(
+            unet=unet,
+            scheduler=scheduler,
+            vae=vae,
+            text_encoder=bert,
+            tokenizer=tokenizer,
+            safety_checker=None,
+            feature_extractor=None,
+            requires_safety_checker=False,
+        )
+        sd_pipe = sd_pipe.to(device)
+        sd_pipe.set_progress_bar_config(disable=None)
+
+        prompt = "A painting of a squirrel eating a burger"
+        generator = torch.Generator(device=device).manual_seed(0)
+        output = sd_pipe([prompt], generator=generator, guidance_scale=6.0, num_inference_steps=2, output_type="np")
+
+        image = output.images
+
+        generator = torch.Generator(device=device).manual_seed(0)
+        image_from_tuple = sd_pipe(
+            [prompt],
+            generator=generator,
+            guidance_scale=6.0,
+            num_inference_steps=2,
+            output_type="np",
+            return_dict=False,
+        )[0]
+
+        image_slice = image[0, -3:, -3:, -1]
+        image_from_tuple_slice = image_from_tuple[0, -3:, -3:, -1]
+
+        assert image.shape == (1, 64, 64, 3)
+        expected_slice = np.array([0.4616, 0.5184, 0.4887, 0.5111, 0.4839, 0.48, 0.5119, 0.5263, 0.4776])
+        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2
+        assert np.abs(image_from_tuple_slice.flatten() - expected_slice).max() < 1e-2
+
+    @unittest.skipIf(torch_device != "cuda", "This test requires a GPU")
+    def test_stable_diffusion_v_pred_fp16(self):
+        """Test that stable diffusion v-prediction works with fp16"""
+        unet = self.dummy_cond_unet
+        scheduler = DDIMScheduler(
+            beta_start=0.00085,
+            beta_end=0.012,
+            beta_schedule="scaled_linear",
+            clip_sample=False,
+            set_alpha_to_one=False,
+            prediction_type="v_prediction",
+        )
+        vae = self.dummy_vae
+        bert = self.dummy_text_encoder
+        tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
+
+        # put models in fp16
+        unet = unet.half()
+        vae = vae.half()
+        bert = bert.half()
+
+        # make sure here that pndm scheduler skips prk
+        sd_pipe = StableDiffusionPipeline(
+            unet=unet,
+            scheduler=scheduler,
+            vae=vae,
+            text_encoder=bert,
+            tokenizer=tokenizer,
+            safety_checker=None,
+            feature_extractor=None,
+            requires_safety_checker=False,
+        )
+        sd_pipe = sd_pipe.to(torch_device)
+        sd_pipe.set_progress_bar_config(disable=None)
+
+        prompt = "A painting of a squirrel eating a burger"
+        generator = torch.Generator(device=torch_device).manual_seed(0)
+        image = sd_pipe([prompt], generator=generator, num_inference_steps=2, output_type="np").images
+
+        assert image.shape == (1, 64, 64, 3)
+
+
+@slow
+@require_torch_gpu
+class StableDiffusion2VPredictionPipelineIntegrationTests(unittest.TestCase):
+    def tearDown(self):
+        # clean up the VRAM after each test
+        super().tearDown()
+        gc.collect()
+        torch.cuda.empty_cache()
+
+    def test_stable_diffusion_v_pred_default(self):
+        sd_pipe = StableDiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-2")
+        sd_pipe = sd_pipe.to(torch_device)
+        sd_pipe.enable_attention_slicing()
+        sd_pipe.set_progress_bar_config(disable=None)
+
+        prompt = "A painting of a squirrel eating a burger"
+        generator = torch.Generator(device=torch_device).manual_seed(0)
+        output = sd_pipe([prompt], generator=generator, guidance_scale=7.5, num_inference_steps=20, output_type="np")
+
+        image = output.images
+        image_slice = image[0, 253:256, 253:256, -1]
+
+        assert image.shape == (1, 768, 768, 3)
+        expected_slice = np.array([0.0567, 0.057, 0.0416, 0.0463, 0.0433, 0.06, 0.0517, 0.0526, 0.0866])
+        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2
+
+    def test_stable_diffusion_v_pred_euler(self):
+        scheduler = EulerDiscreteScheduler.from_pretrained("stabilityai/stable-diffusion-2", subfolder="scheduler")
+        sd_pipe = StableDiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-2", scheduler=scheduler)
+        sd_pipe = sd_pipe.to(torch_device)
+        sd_pipe.enable_attention_slicing()
+        sd_pipe.set_progress_bar_config(disable=None)
+
+        prompt = "A painting of a squirrel eating a burger"
+        generator = torch.Generator(device=torch_device).manual_seed(0)
+
+        output = sd_pipe([prompt], generator=generator, num_inference_steps=5, output_type="numpy")
+        image = output.images
+
+        image_slice = image[0, 253:256, 253:256, -1]
+
+        assert image.shape == (1, 768, 768, 3)
+        expected_slice = np.array([0.0351, 0.0376, 0.0505, 0.0424, 0.0551, 0.0656, 0.0471, 0.0276, 0.0596])
+        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2
+
+    def test_stable_diffusion_v_pred_dpm(self):
+        """
+        TODO: update this test after making DPM compatible with V-prediction!
+        """
+        scheduler = DPMSolverMultistepScheduler.from_pretrained(
+            "stabilityai/stable-diffusion-2", subfolder="scheduler"
+        )
+        sd_pipe = StableDiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-2", scheduler=scheduler)
+        sd_pipe = sd_pipe.to(torch_device)
+        sd_pipe.enable_attention_slicing()
+        sd_pipe.set_progress_bar_config(disable=None)
+
+        prompt = "a photograph of an astronaut riding a horse"
+        generator = torch.Generator(device=torch_device).manual_seed(0)
+        image = sd_pipe(
+            [prompt], generator=generator, guidance_scale=7.5, num_inference_steps=5, output_type="numpy"
+        ).images
+
+        image_slice = image[0, 253:256, 253:256, -1]
+        assert image.shape == (1, 768, 768, 3)
+        expected_slice = np.array([0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0])
+        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2
+
+    def test_stable_diffusion_attention_slicing_v_pred(self):
+        torch.cuda.reset_peak_memory_stats()
+        model_id = "stabilityai/stable-diffusion-2"
+        pipe = StableDiffusionPipeline.from_pretrained(model_id, revision="fp16", torch_dtype=torch.float16)
+        pipe.to(torch_device)
+        pipe.set_progress_bar_config(disable=None)
+
+        prompt = "a photograph of an astronaut riding a horse"
+
+        # make attention efficient
+        pipe.enable_attention_slicing()
+        generator = torch.Generator(device=torch_device).manual_seed(0)
+        with torch.autocast(torch_device):
+            output_chunked = pipe(
+                [prompt], generator=generator, guidance_scale=7.5, num_inference_steps=10, output_type="numpy"
+            )
+            image_chunked = output_chunked.images
+
+        mem_bytes = torch.cuda.max_memory_allocated()
+        torch.cuda.reset_peak_memory_stats()
+        # make sure that less than 5.5 GB is allocated
+        assert mem_bytes < 5.5 * 10**9
+
+        # disable slicing
+        pipe.disable_attention_slicing()
+        generator = torch.Generator(device=torch_device).manual_seed(0)
+        with torch.autocast(torch_device):
+            output = pipe(
+                [prompt], generator=generator, guidance_scale=7.5, num_inference_steps=10, output_type="numpy"
+            )
+            image = output.images
+
+        # make sure that more than 5.5 GB is allocated
+        mem_bytes = torch.cuda.max_memory_allocated()
+        assert mem_bytes > 5.5 * 10**9
+        assert np.abs(image_chunked.flatten() - image.flatten()).max() < 1e-3
+
+    def test_stable_diffusion_text2img_pipeline_v_pred_default(self):
+        expected_image = load_numpy(
+            "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/"
+            "sd2-text2img/astronaut_riding_a_horse_v_pred.npy"
+        )
+
+        pipe = StableDiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-2")
+        pipe.to(torch_device)
+        pipe.enable_attention_slicing()
+        pipe.set_progress_bar_config(disable=None)
+
+        prompt = "astronaut riding a horse"
+
+        generator = torch.Generator(device=torch_device).manual_seed(0)
+        output = pipe(prompt=prompt, guidance_scale=7.5, generator=generator, output_type="np")
+        image = output.images[0]
+
+        assert image.shape == (768, 768, 3)
+        assert np.abs(expected_image - image).max() < 5e-3
+
+    def test_stable_diffusion_text2img_pipeline_v_pred_fp16(self):
+        expected_image = load_numpy(
+            "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/"
+            "sd2-text2img/astronaut_riding_a_horse_v_pred_fp16.npy"
+        )
+
+        pipe = StableDiffusionPipeline.from_pretrained(
+            "stabilityai/stable-diffusion-2", revision="fp16", torch_dtype=torch.float16
+        )
+        pipe.to(torch_device)
+        pipe.set_progress_bar_config(disable=None)
+
+        prompt = "astronaut riding a horse"
+
+        generator = torch.Generator(device=torch_device).manual_seed(0)
+        output = pipe(prompt=prompt, guidance_scale=7.5, generator=generator, output_type="np")
+        image = output.images[0]
+
+        assert image.shape == (768, 768, 3)
+        assert np.abs(expected_image - image).max() < 5e-3
+
+    def test_stable_diffusion_text2img_intermediate_state_v_pred(self):
+        number_of_steps = 0
+
+        def test_callback_fn(step: int, timestep: int, latents: torch.FloatTensor) -> None:
+            test_callback_fn.has_been_called = True
+            nonlocal number_of_steps
+            number_of_steps += 1
+            if step == 0:
+                latents = latents.detach().cpu().numpy()
+                assert latents.shape == (1, 4, 96, 96)
+                latents_slice = latents[0, -3:, -3:, -1]
+                expected_slice = np.array(
+                    [-0.2543, -1.2755, 0.4261, -0.9555, -1.173, -0.5892, 2.4159, 0.1554, -1.2098]
+                )
+                assert np.abs(latents_slice.flatten() - expected_slice).max() < 5e-3
+            elif step == 19:
+                latents = latents.detach().cpu().numpy()
+                assert latents.shape == (1, 4, 96, 96)
+                latents_slice = latents[0, -3:, -3:, -1]
+                expected_slice = np.array(
+                    [-0.9572, -0.967, -0.6152, 0.0894, -0.699, -0.2344, 1.5465, -0.0357, -0.1141]
+                )
+                assert np.abs(latents_slice.flatten() - expected_slice).max() < 1e-2
+
+        test_callback_fn.has_been_called = False
+
+        pipe = StableDiffusionPipeline.from_pretrained(
+            "stabilityai/stable-diffusion-2", revision="fp16", torch_dtype=torch.float16
+        )
+        pipe = pipe.to(torch_device)
+        pipe.set_progress_bar_config(disable=None)
+        pipe.enable_attention_slicing()
+
+        prompt = "Andromeda galaxy in a bottle"
+
+        generator = torch.Generator(device=torch_device).manual_seed(0)
+        with torch.autocast(torch_device):
+            pipe(
+                prompt=prompt,
+                num_inference_steps=20,
+                guidance_scale=7.5,
+                generator=generator,
+                callback=test_callback_fn,
+                callback_steps=1,
+            )
+        assert test_callback_fn.has_been_called
+        assert number_of_steps == 20
+
+    def test_stable_diffusion_low_cpu_mem_usage_v_pred(self):
+        pipeline_id = "stabilityai/stable-diffusion-2"
+
+        start_time = time.time()
+        pipeline_low_cpu_mem_usage = StableDiffusionPipeline.from_pretrained(
+            pipeline_id, revision="fp16", torch_dtype=torch.float16
+        )
+        pipeline_low_cpu_mem_usage.to(torch_device)
+        low_cpu_mem_usage_time = time.time() - start_time
+
+        start_time = time.time()
+        _ = StableDiffusionPipeline.from_pretrained(
+            pipeline_id, revision="fp16", torch_dtype=torch.float16, low_cpu_mem_usage=False
+        )
+        normal_load_time = time.time() - start_time
+
+        assert 2 * low_cpu_mem_usage_time < normal_load_time
+
+    def test_stable_diffusion_pipeline_with_sequential_cpu_offloading_v_pred(self):
+        torch.cuda.empty_cache()
+        torch.cuda.reset_max_memory_allocated()
+        torch.cuda.reset_peak_memory_stats()
+
+        pipeline_id = "stabilityai/stable-diffusion-2"
+        prompt = "Andromeda galaxy in a bottle"
+
+        pipeline = StableDiffusionPipeline.from_pretrained(pipeline_id, revision="fp16", torch_dtype=torch.float16)
+        pipeline = pipeline.to(torch_device)
+        pipeline.enable_attention_slicing(1)
+        pipeline.enable_sequential_cpu_offload()
+
+        generator = torch.Generator(device=torch_device).manual_seed(0)
+        _ = pipeline(prompt, generator=generator, num_inference_steps=5)
+
+        mem_bytes = torch.cuda.max_memory_allocated()
+        # make sure that less than 2.8 GB is allocated
+        assert mem_bytes < 2.8 * 10**9