tests/pipelines/stable_diffusion/test_stable_diffusion_img2img.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,
    LMSDiscreteScheduler,
    PNDMScheduler,
    StableDiffusionImg2ImgPipeline,
    UNet2DConditionModel,
    UNet2DModel,
    VQModel,
)
from diffusers.utils import floats_tensor, load_image, 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 StableDiffusionImg2ImgPipelineFastTests(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_uncond_unet(self):
        torch.manual_seed(0)
        model = UNet2DModel(
            block_out_channels=(32, 64),
            layers_per_block=2,
            sample_size=32,
            in_channels=3,
            out_channels=3,
            down_block_types=("DownBlock2D", "AttnDownBlock2D"),
            up_block_types=("AttnUpBlock2D", "UpBlock2D"),
        )
        return model

    @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,
        )
        return model

    @property
    def dummy_cond_unet_inpaint(self):
        torch.manual_seed(0)
        model = UNet2DConditionModel(
            block_out_channels=(32, 64),
            layers_per_block=2,
            sample_size=32,
            in_channels=9,
            out_channels=4,
            down_block_types=("DownBlock2D", "CrossAttnDownBlock2D"),
            up_block_types=("CrossAttnUpBlock2D", "UpBlock2D"),
            cross_attention_dim=32,
        )
        return model

    @property
    def dummy_vq_model(self):
        torch.manual_seed(0)
        model = VQModel(
            block_out_channels=[32, 64],
            in_channels=3,
            out_channels=3,
            down_block_types=["DownEncoderBlock2D", "DownEncoderBlock2D"],
            up_block_types=["UpDecoderBlock2D", "UpDecoderBlock2D"],
            latent_channels=3,
        )
        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,
        )
        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,
        )
        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_stable_diffusion_img2img_default_case(self):
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
        unet = self.dummy_cond_unet
        scheduler = PNDMScheduler(skip_prk_steps=True)
        vae = self.dummy_vae
        bert = self.dummy_text_encoder
        tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")

        init_image = self.dummy_image.to(device)

        # make sure here that pndm scheduler skips prk
        sd_pipe = StableDiffusionImg2ImgPipeline(
            unet=unet,
            scheduler=scheduler,
            vae=vae,
            text_encoder=bert,
            tokenizer=tokenizer,
            safety_checker=None,
            feature_extractor=self.dummy_extractor,
        )
        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",
            init_image=init_image,
        )

        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",
            init_image=init_image,
            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, 32, 32, 3)
        expected_slice = np.array([0.4492, 0.3865, 0.4222, 0.5854, 0.5139, 0.4379, 0.4193, 0.48, 0.4218])
        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-3
        assert np.abs(image_from_tuple_slice.flatten() - expected_slice).max() < 1e-3

    def test_stable_diffusion_img2img_negative_prompt(self):
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
        unet = self.dummy_cond_unet
        scheduler = PNDMScheduler(skip_prk_steps=True)
        vae = self.dummy_vae
        bert = self.dummy_text_encoder
        tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")

        init_image = self.dummy_image.to(device)

        # make sure here that pndm scheduler skips prk
        sd_pipe = StableDiffusionImg2ImgPipeline(
            unet=unet,
            scheduler=scheduler,
            vae=vae,
            text_encoder=bert,
            tokenizer=tokenizer,
            safety_checker=None,
            feature_extractor=self.dummy_extractor,
        )
        sd_pipe = sd_pipe.to(device)
        sd_pipe.set_progress_bar_config(disable=None)

        prompt = "A painting of a squirrel eating a burger"
        negative_prompt = "french fries"
        generator = torch.Generator(device=device).manual_seed(0)
        output = sd_pipe(
            prompt,
            negative_prompt=negative_prompt,
            generator=generator,
            guidance_scale=6.0,
            num_inference_steps=2,
            output_type="np",
            init_image=init_image,
        )
        image = output.images
        image_slice = image[0, -3:, -3:, -1]

        assert image.shape == (1, 32, 32, 3)
        expected_slice = np.array([0.4065, 0.3783, 0.4050, 0.5266, 0.4781, 0.4252, 0.4203, 0.4692, 0.4365])
        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-3

    def test_stable_diffusion_img2img_multiple_init_images(self):
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
        unet = self.dummy_cond_unet
        scheduler = PNDMScheduler(skip_prk_steps=True)
        vae = self.dummy_vae
        bert = self.dummy_text_encoder
        tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")

        init_image = self.dummy_image.to(device).repeat(2, 1, 1, 1)

        # make sure here that pndm scheduler skips prk
        sd_pipe = StableDiffusionImg2ImgPipeline(
            unet=unet,
            scheduler=scheduler,
            vae=vae,
            text_encoder=bert,
            tokenizer=tokenizer,
            safety_checker=None,
            feature_extractor=self.dummy_extractor,
        )
        sd_pipe = sd_pipe.to(device)
        sd_pipe.set_progress_bar_config(disable=None)

        prompt = 2 * ["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",
            init_image=init_image,
        )

        image = output.images

        image_slice = image[-1, -3:, -3:, -1]

        assert image.shape == (2, 32, 32, 3)
        expected_slice = np.array([0.5144, 0.4447, 0.4735, 0.6676, 0.5526, 0.5454, 0.645, 0.5149, 0.4689])
        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-3

    def test_stable_diffusion_img2img_k_lms(self):
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
        unet = self.dummy_cond_unet
        scheduler = LMSDiscreteScheduler(beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear")

        vae = self.dummy_vae
        bert = self.dummy_text_encoder
        tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")

        init_image = self.dummy_image.to(device)

        # make sure here that pndm scheduler skips prk
        sd_pipe = StableDiffusionImg2ImgPipeline(
            unet=unet,
            scheduler=scheduler,
            vae=vae,
            text_encoder=bert,
            tokenizer=tokenizer,
            safety_checker=None,
            feature_extractor=self.dummy_extractor,
        )
        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",
            init_image=init_image,
        )
        image = output.images

        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",
            init_image=init_image,
            return_dict=False,
        )
        image_from_tuple = output[0]

        image_slice = image[0, -3:, -3:, -1]
        image_from_tuple_slice = image_from_tuple[0, -3:, -3:, -1]

        assert image.shape == (1, 32, 32, 3)
        expected_slice = np.array([0.4367, 0.4986, 0.4372, 0.6706, 0.5665, 0.444, 0.5864, 0.6019, 0.5203])
        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-3
        assert np.abs(image_from_tuple_slice.flatten() - expected_slice).max() < 1e-3

    def test_stable_diffusion_img2img_num_images_per_prompt(self):
        device = "cpu"
        unet = self.dummy_cond_unet
        scheduler = PNDMScheduler(skip_prk_steps=True)
        vae = self.dummy_vae
        bert = self.dummy_text_encoder
        tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")

        init_image = self.dummy_image.to(device)

        # make sure here that pndm scheduler skips prk
        sd_pipe = StableDiffusionImg2ImgPipeline(
            unet=unet,
            scheduler=scheduler,
            vae=vae,
            text_encoder=bert,
            tokenizer=tokenizer,
            safety_checker=None,
            feature_extractor=self.dummy_extractor,
        )
        sd_pipe = sd_pipe.to(device)
        sd_pipe.set_progress_bar_config(disable=None)

        prompt = "A painting of a squirrel eating a burger"

        # test num_images_per_prompt=1 (default)
        images = sd_pipe(
            prompt,
            num_inference_steps=2,
            output_type="np",
            init_image=init_image,
        ).images

        assert images.shape == (1, 32, 32, 3)

        # test num_images_per_prompt=1 (default) for batch of prompts
        batch_size = 2
        images = sd_pipe(
            [prompt] * batch_size,
            num_inference_steps=2,
            output_type="np",
            init_image=init_image,
        ).images

        assert images.shape == (batch_size, 32, 32, 3)

        # test num_images_per_prompt for single prompt
        num_images_per_prompt = 2
        images = sd_pipe(
            prompt,
            num_inference_steps=2,
            output_type="np",
            init_image=init_image,
            num_images_per_prompt=num_images_per_prompt,
        ).images

        assert images.shape == (num_images_per_prompt, 32, 32, 3)

        # test num_images_per_prompt for batch of prompts
        batch_size = 2
        images = sd_pipe(
            [prompt] * batch_size,
            num_inference_steps=2,
            output_type="np",
            init_image=init_image,
            num_images_per_prompt=num_images_per_prompt,
        ).images

        assert images.shape == (batch_size * num_images_per_prompt, 32, 32, 3)

    @unittest.skipIf(torch_device != "cuda", "This test requires a GPU")
    def test_stable_diffusion_img2img_fp16(self):
        """Test that stable diffusion img2img works with fp16"""
        unet = self.dummy_cond_unet
        scheduler = PNDMScheduler(skip_prk_steps=True)
        vae = self.dummy_vae
        bert = self.dummy_text_encoder
        tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")

        init_image = self.dummy_image.to(torch_device)

        # put models in fp16
        unet = unet.half()
        vae = vae.half()
        bert = bert.half()

        # make sure here that pndm scheduler skips prk
        sd_pipe = StableDiffusionImg2ImgPipeline(
            unet=unet,
            scheduler=scheduler,
            vae=vae,
            text_encoder=bert,
            tokenizer=tokenizer,
            safety_checker=None,
            feature_extractor=self.dummy_extractor,
        )
        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",
            init_image=init_image,
        ).images

        assert image.shape == (1, 32, 32, 3)


@slow
@require_torch_gpu
class StableDiffusionImg2ImgPipelineIntegrationTests(unittest.TestCase):
    def tearDown(self):
        # clean up the VRAM after each test
        super().tearDown()
        gc.collect()
        torch.cuda.empty_cache()

    def test_stable_diffusion_img2img_pipeline_default(self):
        init_image = load_image(
            "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main"
            "/img2img/sketch-mountains-input.jpg"
        )
        init_image = init_image.resize((768, 512))
        expected_image = load_numpy(
            "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/img2img/fantasy_landscape.npy"
        )

        model_id = "CompVis/stable-diffusion-v1-4"
        pipe = StableDiffusionImg2ImgPipeline.from_pretrained(
            model_id,
            safety_checker=None,
        )
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        pipe.enable_attention_slicing()

        prompt = "A fantasy landscape, trending on artstation"

        generator = torch.Generator(device=torch_device).manual_seed(0)
        output = pipe(
            prompt=prompt,
            init_image=init_image,
            strength=0.75,
            guidance_scale=7.5,
            generator=generator,
            output_type="np",
        )
        image = output.images[0]

        assert image.shape == (512, 768, 3)
        # img2img is flaky across GPUs even in fp32, so using MAE here
        assert np.abs(expected_image - image).max() < 1e-3

    def test_stable_diffusion_img2img_pipeline_k_lms(self):
        init_image = load_image(
            "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main"
            "/img2img/sketch-mountains-input.jpg"
        )
        init_image = init_image.resize((768, 512))
        expected_image = load_numpy(
            "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/img2img/fantasy_landscape_k_lms.npy"
        )

        model_id = "CompVis/stable-diffusion-v1-4"
        lms = LMSDiscreteScheduler.from_pretrained(model_id, subfolder="scheduler")
        pipe = StableDiffusionImg2ImgPipeline.from_pretrained(
            model_id,
            scheduler=lms,
            safety_checker=None,
        )
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        pipe.enable_attention_slicing()

        prompt = "A fantasy landscape, trending on artstation"

        generator = torch.Generator(device=torch_device).manual_seed(0)
        output = pipe(
            prompt=prompt,
            init_image=init_image,
            strength=0.75,
            guidance_scale=7.5,
            generator=generator,
            output_type="np",
        )
        image = output.images[0]

        assert image.shape == (512, 768, 3)
        assert np.abs(expected_image - image).max() < 1e-3

    def test_stable_diffusion_img2img_pipeline_ddim(self):
        init_image = load_image(
            "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main"
            "/img2img/sketch-mountains-input.jpg"
        )
        init_image = init_image.resize((768, 512))
        expected_image = load_numpy(
            "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/img2img/fantasy_landscape_ddim.npy"
        )

        model_id = "CompVis/stable-diffusion-v1-4"
        ddim = DDIMScheduler.from_pretrained(model_id, subfolder="scheduler")
        pipe = StableDiffusionImg2ImgPipeline.from_pretrained(
            model_id,
            scheduler=ddim,
            safety_checker=None,
        )
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        pipe.enable_attention_slicing()

        prompt = "A fantasy landscape, trending on artstation"

        generator = torch.Generator(device=torch_device).manual_seed(0)
        output = pipe(
            prompt=prompt,
            init_image=init_image,
            strength=0.75,
            guidance_scale=7.5,
            generator=generator,
            output_type="np",
        )
        image = output.images[0]

        assert image.shape == (512, 768, 3)
        assert np.abs(expected_image - image).max() < 1e-3

    def test_stable_diffusion_img2img_intermediate_state(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, 64, 96)
                latents_slice = latents[0, -3:, -3:, -1]
                expected_slice = np.array([0.9052, -0.0184, 0.4810, 0.2898, 0.5851, 1.4920, 0.5362, 1.9838, 0.0530])
                assert np.abs(latents_slice.flatten() - expected_slice).max() < 1e-3
            elif step == 37:
                latents = latents.detach().cpu().numpy()
                assert latents.shape == (1, 4, 64, 96)
                latents_slice = latents[0, -3:, -3:, -1]
                expected_slice = np.array([0.7071, 0.7831, 0.8300, 1.8140, 1.7840, 1.9402, 1.3651, 1.6590, 1.2828])
                assert np.abs(latents_slice.flatten() - expected_slice).max() < 1e-2

        test_callback_fn.has_been_called = False

        init_image = load_image(
            "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main"
            "/img2img/sketch-mountains-input.jpg"
        )
        init_image = init_image.resize((768, 512))

        pipe = StableDiffusionImg2ImgPipeline.from_pretrained(
            "CompVis/stable-diffusion-v1-4",
            revision="fp16",
            torch_dtype=torch.float16,
        )
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        pipe.enable_attention_slicing()

        prompt = "A fantasy landscape, trending on artstation"

        generator = torch.Generator(device=torch_device).manual_seed(0)
        with torch.autocast(torch_device):
            pipe(
                prompt=prompt,
                init_image=init_image,
                strength=0.75,
                num_inference_steps=50,
                guidance_scale=7.5,
                generator=generator,
                callback=test_callback_fn,
                callback_steps=1,
            )
        assert test_callback_fn.has_been_called
        assert number_of_steps == 38

    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()

        init_image = load_image(
            "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main"
            "/img2img/sketch-mountains-input.jpg"
        )
        init_image = init_image.resize((768, 512))

        model_id = "CompVis/stable-diffusion-v1-4"
        lms = LMSDiscreteScheduler.from_pretrained(model_id, subfolder="scheduler")
        pipe = StableDiffusionImg2ImgPipeline.from_pretrained(
            model_id, scheduler=lms, safety_checker=None, device_map="auto", 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 fantasy landscape, trending on artstation"

        generator = torch.Generator(device=torch_device).manual_seed(0)
        _ = pipe(
            prompt=prompt,
            init_image=init_image,
            strength=0.75,
            guidance_scale=7.5,
            generator=generator,
            output_type="np",
            num_inference_steps=5,
        )

        mem_bytes = torch.cuda.max_memory_allocated()
        # make sure that less than 2.2 GB is allocated
        assert mem_bytes < 2.2 * 10**9