diffusers/tests/pipelines/controlnet/test_controlnet_inpaint_sdxl.py

# coding=utf-8
# Copyright 2024 Harutatsu Akiyama, Jinbin Bai, 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 random
import unittest

import numpy as np
import torch
from PIL import Image
from transformers import (
    CLIPImageProcessor,
    CLIPTextConfig,
    CLIPTextModel,
    CLIPTextModelWithProjection,
    CLIPTokenizer,
    CLIPVisionConfig,
    CLIPVisionModelWithProjection,
)

from diffusers import (
    AutoencoderKL,
    ControlNetModel,
    EulerDiscreteScheduler,
    StableDiffusionXLControlNetInpaintPipeline,
    UNet2DConditionModel,
)
from diffusers.utils.import_utils import is_xformers_available
from diffusers.utils.testing_utils import (
    enable_full_determinism,
    floats_tensor,
    require_torch_gpu,
    torch_device,
)

from ..pipeline_params import (
    IMAGE_TO_IMAGE_IMAGE_PARAMS,
    TEXT_TO_IMAGE_BATCH_PARAMS,
    TEXT_TO_IMAGE_CALLBACK_CFG_PARAMS,
    TEXT_TO_IMAGE_IMAGE_PARAMS,
    TEXT_TO_IMAGE_PARAMS,
)
from ..test_pipelines_common import (
    PipelineKarrasSchedulerTesterMixin,
    PipelineLatentTesterMixin,
    PipelineTesterMixin,
)


enable_full_determinism()


class ControlNetPipelineSDXLFastTests(
    PipelineLatentTesterMixin, PipelineKarrasSchedulerTesterMixin, PipelineTesterMixin, unittest.TestCase
):
    pipeline_class = StableDiffusionXLControlNetInpaintPipeline
    params = TEXT_TO_IMAGE_PARAMS
    batch_params = TEXT_TO_IMAGE_BATCH_PARAMS
    image_params = frozenset(IMAGE_TO_IMAGE_IMAGE_PARAMS.union({"mask_image", "control_image"}))
    image_latents_params = TEXT_TO_IMAGE_IMAGE_PARAMS
    callback_cfg_params = TEXT_TO_IMAGE_CALLBACK_CFG_PARAMS.union(
        {
            "add_text_embeds",
            "add_time_ids",
            "mask",
            "masked_image_latents",
        }
    )

    supports_dduf = False

    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=4,
            out_channels=4,
            down_block_types=("DownBlock2D", "CrossAttnDownBlock2D"),
            up_block_types=("CrossAttnUpBlock2D", "UpBlock2D"),
            # SD2-specific config below
            attention_head_dim=(2, 4),
            use_linear_projection=True,
            addition_embed_type="text_time",
            addition_time_embed_dim=8,
            transformer_layers_per_block=(1, 2),
            projection_class_embeddings_input_dim=80,  # 6 * 8 + 32
            cross_attention_dim=64,
        )
        torch.manual_seed(0)
        controlnet = ControlNetModel(
            block_out_channels=(32, 64),
            layers_per_block=2,
            in_channels=4,
            down_block_types=("DownBlock2D", "CrossAttnDownBlock2D"),
            conditioning_embedding_out_channels=(16, 32),
            # SD2-specific config below
            attention_head_dim=(2, 4),
            use_linear_projection=True,
            addition_embed_type="text_time",
            addition_time_embed_dim=8,
            transformer_layers_per_block=(1, 2),
            projection_class_embeddings_input_dim=80,  # 6 * 8 + 32
            cross_attention_dim=64,
        )
        scheduler = EulerDiscreteScheduler(
            beta_start=0.00085,
            beta_end=0.012,
            steps_offset=1,
            beta_schedule="scaled_linear",
            timestep_spacing="leading",
        )
        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,
        )
        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=32,
        )
        text_encoder = CLIPTextModel(text_encoder_config)
        tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")

        torch.manual_seed(0)
        text_encoder_2 = CLIPTextModelWithProjection(text_encoder_config)
        tokenizer_2 = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")

        image_encoder_config = CLIPVisionConfig(
            hidden_size=32,
            image_size=224,
            projection_dim=32,
            intermediate_size=37,
            num_attention_heads=4,
            num_channels=3,
            num_hidden_layers=5,
            patch_size=14,
        )

        image_encoder = CLIPVisionModelWithProjection(image_encoder_config)

        feature_extractor = CLIPImageProcessor(
            crop_size=224,
            do_center_crop=True,
            do_normalize=True,
            do_resize=True,
            image_mean=[0.48145466, 0.4578275, 0.40821073],
            image_std=[0.26862954, 0.26130258, 0.27577711],
            resample=3,
            size=224,
        )

        components = {
            "unet": unet,
            "controlnet": controlnet,
            "scheduler": scheduler,
            "vae": vae,
            "text_encoder": text_encoder,
            "tokenizer": tokenizer,
            "text_encoder_2": text_encoder_2,
            "tokenizer_2": tokenizer_2,
            "image_encoder": image_encoder,
            "feature_extractor": feature_extractor,
        }
        return components

    def get_dummy_inputs(self, device, seed=0, img_res=64):
        if str(device).startswith("mps"):
            generator = torch.manual_seed(seed)
        else:
            generator = torch.Generator(device=device).manual_seed(seed)

        # Get random floats in [0, 1] as image
        image = floats_tensor((1, 3, 32, 32), rng=random.Random(seed)).to(device)
        image = image.cpu().permute(0, 2, 3, 1)[0]
        mask_image = torch.ones_like(image)
        controlnet_embedder_scale_factor = 2
        control_image = (
            floats_tensor(
                (1, 3, 32 * controlnet_embedder_scale_factor, 32 * controlnet_embedder_scale_factor),
                rng=random.Random(seed),
            )
            .to(device)
            .cpu()
        )
        control_image = control_image.cpu().permute(0, 2, 3, 1)[0]
        # Convert image and mask_image to [0, 255]
        image = 255 * image
        mask_image = 255 * mask_image
        control_image = 255 * control_image
        # Convert to PIL image
        init_image = Image.fromarray(np.uint8(image)).convert("RGB").resize((img_res, img_res))
        mask_image = Image.fromarray(np.uint8(mask_image)).convert("L").resize((img_res, img_res))
        control_image = Image.fromarray(np.uint8(control_image)).convert("RGB").resize((img_res, img_res))

        inputs = {
            "prompt": "A painting of a squirrel eating a burger",
            "generator": generator,
            "num_inference_steps": 2,
            "guidance_scale": 6.0,
            "output_type": "np",
            "image": init_image,
            "mask_image": mask_image,
            "control_image": control_image,
        }
        return inputs

    def test_attention_slicing_forward_pass(self):
        return self._test_attention_slicing_forward_pass(expected_max_diff=2e-3)

    @unittest.skipIf(
        torch_device != "cuda" or not is_xformers_available(),
        reason="XFormers attention is only available with CUDA and `xformers` installed",
    )
    def test_xformers_attention_forwardGenerator_pass(self):
        self._test_xformers_attention_forwardGenerator_pass(expected_max_diff=2e-3)

    def test_inference_batch_single_identical(self):
        self._test_inference_batch_single_identical(expected_max_diff=2e-3)

    @require_torch_gpu
    def test_stable_diffusion_xl_offloads(self):
        pipes = []
        components = self.get_dummy_components()
        sd_pipe = self.pipeline_class(**components).to(torch_device)
        pipes.append(sd_pipe)

        components = self.get_dummy_components()
        sd_pipe = self.pipeline_class(**components)
        sd_pipe.enable_model_cpu_offload()
        pipes.append(sd_pipe)

        components = self.get_dummy_components()
        sd_pipe = self.pipeline_class(**components)
        sd_pipe.enable_sequential_cpu_offload()
        pipes.append(sd_pipe)

        image_slices = []
        for pipe in pipes:
            pipe.unet.set_default_attn_processor()

            inputs = self.get_dummy_inputs(torch_device)
            image = pipe(**inputs).images

            image_slices.append(image[0, -3:, -3:, -1].flatten())

        assert np.abs(image_slices[0] - image_slices[1]).max() < 1e-3
        assert np.abs(image_slices[0] - image_slices[2]).max() < 1e-3

    def test_stable_diffusion_xl_multi_prompts(self):
        components = self.get_dummy_components()
        sd_pipe = self.pipeline_class(**components).to(torch_device)

        # forward with single prompt
        inputs = self.get_dummy_inputs(torch_device)
        output = sd_pipe(**inputs)
        image_slice_1 = output.images[0, -3:, -3:, -1]

        # forward with same prompt duplicated
        inputs = self.get_dummy_inputs(torch_device)
        inputs["prompt_2"] = inputs["prompt"]
        output = sd_pipe(**inputs)
        image_slice_2 = output.images[0, -3:, -3:, -1]

        # ensure the results are equal
        assert np.abs(image_slice_1.flatten() - image_slice_2.flatten()).max() < 1e-4

        # forward with different prompt
        inputs = self.get_dummy_inputs(torch_device)
        inputs["prompt_2"] = "different prompt"
        output = sd_pipe(**inputs)
        image_slice_3 = output.images[0, -3:, -3:, -1]

        # ensure the results are not equal
        assert np.abs(image_slice_1.flatten() - image_slice_3.flatten()).max() > 1e-4

        # manually set a negative_prompt
        inputs = self.get_dummy_inputs(torch_device)
        inputs["negative_prompt"] = "negative prompt"
        output = sd_pipe(**inputs)
        image_slice_1 = output.images[0, -3:, -3:, -1]

        # forward with same negative_prompt duplicated
        inputs = self.get_dummy_inputs(torch_device)
        inputs["negative_prompt"] = "negative prompt"
        inputs["negative_prompt_2"] = inputs["negative_prompt"]
        output = sd_pipe(**inputs)
        image_slice_2 = output.images[0, -3:, -3:, -1]

        # ensure the results are equal
        assert np.abs(image_slice_1.flatten() - image_slice_2.flatten()).max() < 1e-4

        # forward with different negative_prompt
        inputs = self.get_dummy_inputs(torch_device)
        inputs["negative_prompt"] = "negative prompt"
        inputs["negative_prompt_2"] = "different negative prompt"
        output = sd_pipe(**inputs)
        image_slice_3 = output.images[0, -3:, -3:, -1]

        # ensure the results are not equal
        assert np.abs(image_slice_1.flatten() - image_slice_3.flatten()).max() > 1e-4

    def test_controlnet_sdxl_guess(self):
        device = "cpu"

        components = self.get_dummy_components()

        sd_pipe = self.pipeline_class(**components)
        sd_pipe = sd_pipe.to(device)

        sd_pipe.set_progress_bar_config(disable=None)

        inputs = self.get_dummy_inputs(device)
        inputs["guess_mode"] = True

        output = sd_pipe(**inputs)
        image_slice = output.images[0, -3:, -3:, -1]

        expected_slice = np.array([0.5460, 0.4943, 0.4635, 0.5832, 0.5366, 0.4815, 0.6034, 0.5741, 0.4341])

        # make sure that it's equal
        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-4

    # TODO(Patrick, Sayak) - skip for now as this requires more refiner tests
    def test_save_load_optional_components(self):
        pass

    def test_float16_inference(self):
        super().test_float16_inference(expected_max_diff=5e-1)