diffusers/tests/pipelines/stable_diffusion_2/test_stable_diffusion_inpaint.py

# coding=utf-8
# Copyright 2025 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 PIL import Image
from transformers import CLIPTextConfig, CLIPTextModel, CLIPTokenizer

from diffusers import AutoencoderKL, PNDMScheduler, StableDiffusionInpaintPipeline, UNet2DConditionModel
from diffusers.utils.testing_utils import (
    backend_empty_cache,
    backend_max_memory_allocated,
    backend_reset_max_memory_allocated,
    backend_reset_peak_memory_stats,
    enable_full_determinism,
    floats_tensor,
    load_image,
    load_numpy,
    require_torch_accelerator,
    slow,
    torch_device,
)

from ..pipeline_params import (
    TEXT_GUIDED_IMAGE_INPAINTING_BATCH_PARAMS,
    TEXT_GUIDED_IMAGE_INPAINTING_PARAMS,
    TEXT_TO_IMAGE_CALLBACK_CFG_PARAMS,
)
from ..test_pipelines_common import PipelineKarrasSchedulerTesterMixin, PipelineLatentTesterMixin, PipelineTesterMixin


enable_full_determinism()


class StableDiffusion2InpaintPipelineFastTests(
    PipelineLatentTesterMixin, PipelineKarrasSchedulerTesterMixin, PipelineTesterMixin, unittest.TestCase
):
    pipeline_class = StableDiffusionInpaintPipeline
    params = TEXT_GUIDED_IMAGE_INPAINTING_PARAMS
    batch_params = TEXT_GUIDED_IMAGE_INPAINTING_BATCH_PARAMS
    image_params = frozenset(
        []
    )  # TO-DO: update image_params once pipeline is refactored with VaeImageProcessor.preprocess
    image_latents_params = frozenset([])
    callback_cfg_params = TEXT_TO_IMAGE_CALLBACK_CFG_PARAMS.union({"mask", "masked_image_latents"})

    def get_dummy_components(self):
        torch.manual_seed(0)
        unet = 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,
            # SD2-specific config below
            attention_head_dim=(2, 4),
            use_linear_projection=True,
        )
        scheduler = PNDMScheduler(skip_prk_steps=True)
        torch.manual_seed(0)
        vae = 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,
        )
        torch.manual_seed(0)
        text_encoder_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,
        )
        text_encoder = CLIPTextModel(text_encoder_config)
        tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")

        components = {
            "unet": unet,
            "scheduler": scheduler,
            "vae": vae,
            "text_encoder": text_encoder,
            "tokenizer": tokenizer,
            "safety_checker": None,
            "feature_extractor": None,
            "image_encoder": None,
        }
        return components

    def get_dummy_inputs(self, device, seed=0):
        # TODO: use tensor inputs instead of PIL, this is here just to leave the old expected_slices untouched
        image = floats_tensor((1, 3, 32, 32), rng=random.Random(seed)).to(device)
        image = image.cpu().permute(0, 2, 3, 1)[0]
        init_image = Image.fromarray(np.uint8(image)).convert("RGB").resize((64, 64))
        mask_image = Image.fromarray(np.uint8(image + 4)).convert("RGB").resize((64, 64))
        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",
            "image": init_image,
            "mask_image": mask_image,
            "generator": generator,
            "num_inference_steps": 2,
            "guidance_scale": 6.0,
            "output_type": "np",
        }
        return inputs

    def test_stable_diffusion_inpaint(self):
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
        components = self.get_dummy_components()
        sd_pipe = StableDiffusionInpaintPipeline(**components)
        sd_pipe = sd_pipe.to(device)
        sd_pipe.set_progress_bar_config(disable=None)

        inputs = self.get_dummy_inputs(device)
        image = sd_pipe(**inputs).images
        image_slice = image[0, -3:, -3:, -1]

        assert image.shape == (1, 64, 64, 3)
        expected_slice = np.array([0.4727, 0.5735, 0.3941, 0.5446, 0.5926, 0.4394, 0.5062, 0.4654, 0.4476])

        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2

    def test_inference_batch_single_identical(self):
        super().test_inference_batch_single_identical(expected_max_diff=3e-3)

    def test_encode_prompt_works_in_isolation(self):
        extra_required_param_value_dict = {
            "device": torch.device(torch_device).type,
            "do_classifier_free_guidance": self.get_dummy_inputs(device=torch_device).get("guidance_scale", 1.0) > 1.0,
        }
        return super().test_encode_prompt_works_in_isolation(extra_required_param_value_dict)


@slow
@require_torch_accelerator
class StableDiffusionInpaintPipelineIntegrationTests(unittest.TestCase):
    def setUp(self):
        # clean up the VRAM before each test
        super().setUp()
        gc.collect()
        backend_empty_cache(torch_device)

    def tearDown(self):
        # clean up the VRAM after each test
        super().tearDown()
        gc.collect()
        backend_empty_cache(torch_device)

    def test_stable_diffusion_inpaint_pipeline(self):
        init_image = load_image(
            "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main"
            "/sd2-inpaint/init_image.png"
        )
        mask_image = load_image(
            "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/sd2-inpaint/mask.png"
        )
        expected_image = load_numpy(
            "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/sd2-inpaint"
            "/yellow_cat_sitting_on_a_park_bench.npy"
        )

        model_id = "stabilityai/stable-diffusion-2-inpainting"
        pipe = StableDiffusionInpaintPipeline.from_pretrained(model_id, safety_checker=None)
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        pipe.enable_attention_slicing()

        prompt = "Face of a yellow cat, high resolution, sitting on a park bench"

        generator = torch.manual_seed(0)
        output = pipe(
            prompt=prompt,
            image=init_image,
            mask_image=mask_image,
            generator=generator,
            output_type="np",
        )
        image = output.images[0]

        assert image.shape == (512, 512, 3)
        assert np.abs(expected_image - image).max() < 9e-3

    def test_stable_diffusion_inpaint_pipeline_fp16(self):
        init_image = load_image(
            "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main"
            "/sd2-inpaint/init_image.png"
        )
        mask_image = load_image(
            "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/sd2-inpaint/mask.png"
        )
        expected_image = load_numpy(
            "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/sd2-inpaint"
            "/yellow_cat_sitting_on_a_park_bench_fp16.npy"
        )

        model_id = "stabilityai/stable-diffusion-2-inpainting"
        pipe = StableDiffusionInpaintPipeline.from_pretrained(
            model_id,
            torch_dtype=torch.float16,
            safety_checker=None,
        )
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        pipe.enable_attention_slicing()

        prompt = "Face of a yellow cat, high resolution, sitting on a park bench"

        generator = torch.manual_seed(0)
        output = pipe(
            prompt=prompt,
            image=init_image,
            mask_image=mask_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):
        backend_empty_cache(torch_device)
        backend_reset_max_memory_allocated(torch_device)
        backend_reset_peak_memory_stats(torch_device)

        init_image = load_image(
            "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main"
            "/sd2-inpaint/init_image.png"
        )
        mask_image = load_image(
            "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/sd2-inpaint/mask.png"
        )

        model_id = "stabilityai/stable-diffusion-2-inpainting"
        pndm = PNDMScheduler.from_pretrained(model_id, subfolder="scheduler")
        pipe = StableDiffusionInpaintPipeline.from_pretrained(
            model_id,
            safety_checker=None,
            scheduler=pndm,
            torch_dtype=torch.float16,
        )
        pipe.set_progress_bar_config(disable=None)
        pipe.enable_attention_slicing(1)
        pipe.enable_sequential_cpu_offload(device=torch_device)

        prompt = "Face of a yellow cat, high resolution, sitting on a park bench"

        generator = torch.manual_seed(0)
        _ = pipe(
            prompt=prompt,
            image=init_image,
            mask_image=mask_image,
            generator=generator,
            num_inference_steps=2,
            output_type="np",
        )

        mem_bytes = backend_max_memory_allocated(torch_device)
        # make sure that less than 2.65 GB is allocated
        assert mem_bytes < 2.65 * 10**9