import gc import tempfile import torch from diffusers import ControlNetModel, StableDiffusionXLControlNetPipeline from diffusers.loaders.single_file_utils import _extract_repo_id_and_weights_name from diffusers.utils import load_image from ..testing_utils import ( backend_empty_cache, enable_full_determinism, numpy_cosine_similarity_distance, require_torch_accelerator, slow, torch_device, ) from .single_file_testing_utils import ( SDXLSingleFileTesterMixin, download_diffusers_config, download_single_file_checkpoint, ) enable_full_determinism() @slow @require_torch_accelerator class TestStableDiffusionXLControlNetPipelineSingleFileSlow(SDXLSingleFileTesterMixin): pipeline_class = StableDiffusionXLControlNetPipeline ckpt_path = "https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/blob/main/sd_xl_base_1.0.safetensors" repo_id = "stabilityai/stable-diffusion-xl-base-1.0" original_config = ( "https://raw.githubusercontent.com/Stability-AI/generative-models/main/configs/inference/sd_xl_base.yaml" ) def setup_method(self): gc.collect() backend_empty_cache(torch_device) def teardown_method(self): gc.collect() backend_empty_cache(torch_device) def get_inputs(self, device, generator_device="cpu", dtype=torch.float32, seed=0): generator = torch.Generator(device=generator_device).manual_seed(seed) image = load_image( "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/sd_controlnet/stormtrooper_depth.png" ) inputs = { "prompt": "Stormtrooper's lecture", "image": image, "generator": generator, "num_inference_steps": 2, "strength": 0.75, "guidance_scale": 7.5, "output_type": "np", } return inputs def test_single_file_format_inference_is_same_as_pretrained(self): controlnet = ControlNetModel.from_pretrained("diffusers/controlnet-depth-sdxl-1.0", torch_dtype=torch.float16) pipe_single_file = self.pipeline_class.from_single_file( self.ckpt_path, controlnet=controlnet, torch_dtype=torch.float16 ) pipe_single_file.unet.set_default_attn_processor() pipe_single_file.enable_model_cpu_offload(device=torch_device) pipe_single_file.set_progress_bar_config(disable=None) inputs = self.get_inputs(torch_device) single_file_images = pipe_single_file(**inputs).images[0] pipe = self.pipeline_class.from_pretrained(self.repo_id, controlnet=controlnet, torch_dtype=torch.float16) pipe.unet.set_default_attn_processor() pipe.enable_model_cpu_offload(device=torch_device) inputs = self.get_inputs(torch_device) images = pipe(**inputs).images[0] assert images.shape == (512, 512, 3) assert single_file_images.shape == (512, 512, 3) max_diff = numpy_cosine_similarity_distance(images[0].flatten(), single_file_images[0].flatten()) assert max_diff < 5e-2 def test_single_file_components(self): controlnet = ControlNetModel.from_pretrained( "diffusers/controlnet-depth-sdxl-1.0", torch_dtype=torch.float16, variant="fp16" ) pipe = self.pipeline_class.from_pretrained( self.repo_id, variant="fp16", controlnet=controlnet, torch_dtype=torch.float16, ) pipe_single_file = self.pipeline_class.from_single_file(self.ckpt_path, controlnet=controlnet) super().test_single_file_components(pipe, pipe_single_file) def test_single_file_components_local_files_only(self): controlnet = ControlNetModel.from_pretrained( "diffusers/controlnet-depth-sdxl-1.0", torch_dtype=torch.float16, variant="fp16" ) pipe = self.pipeline_class.from_pretrained( self.repo_id, variant="fp16", controlnet=controlnet, torch_dtype=torch.float16, ) with tempfile.TemporaryDirectory() as tmpdir: repo_id, weight_name = _extract_repo_id_and_weights_name(self.ckpt_path) local_ckpt_path = download_single_file_checkpoint(repo_id, weight_name, tmpdir) single_file_pipe = self.pipeline_class.from_single_file( local_ckpt_path, controlnet=controlnet, safety_checker=None, local_files_only=True ) self._compare_component_configs(pipe, single_file_pipe) def test_single_file_components_with_original_config(self): controlnet = ControlNetModel.from_pretrained( "diffusers/controlnet-depth-sdxl-1.0", torch_dtype=torch.float16, variant="fp16" ) pipe = self.pipeline_class.from_pretrained( self.repo_id, variant="fp16", controlnet=controlnet, torch_dtype=torch.float16, ) pipe_single_file = self.pipeline_class.from_single_file( self.ckpt_path, original_config=self.original_config, controlnet=controlnet, ) self._compare_component_configs(pipe, pipe_single_file) def test_single_file_components_with_original_config_local_files_only(self): controlnet = ControlNetModel.from_pretrained( "diffusers/controlnet-depth-sdxl-1.0", torch_dtype=torch.float16, variant="fp16" ) pipe = self.pipeline_class.from_pretrained( self.repo_id, variant="fp16", controlnet=controlnet, torch_dtype=torch.float16, ) with tempfile.TemporaryDirectory() as tmpdir: repo_id, weight_name = _extract_repo_id_and_weights_name(self.ckpt_path) local_ckpt_path = download_single_file_checkpoint(repo_id, weight_name, tmpdir) pipe_single_file = self.pipeline_class.from_single_file( local_ckpt_path, safety_checker=None, controlnet=controlnet, local_files_only=True, ) self._compare_component_configs(pipe, pipe_single_file) def test_single_file_components_with_diffusers_config(self): controlnet = ControlNetModel.from_pretrained( "diffusers/controlnet-depth-sdxl-1.0", torch_dtype=torch.float16, variant="fp16" ) pipe = self.pipeline_class.from_pretrained(self.repo_id, controlnet=controlnet) pipe_single_file = self.pipeline_class.from_single_file( self.ckpt_path, controlnet=controlnet, config=self.repo_id ) super()._compare_component_configs(pipe, pipe_single_file) def test_single_file_components_with_diffusers_config_local_files_only(self): controlnet = ControlNetModel.from_pretrained( "diffusers/controlnet-depth-sdxl-1.0", torch_dtype=torch.float16, variant="fp16" ) pipe = self.pipeline_class.from_pretrained( self.repo_id, controlnet=controlnet, ) with tempfile.TemporaryDirectory() as tmpdir: repo_id, weight_name = _extract_repo_id_and_weights_name(self.ckpt_path) local_ckpt_path = download_single_file_checkpoint(repo_id, weight_name, tmpdir) local_diffusers_config = download_diffusers_config(self.repo_id, tmpdir) pipe_single_file = self.pipeline_class.from_single_file( local_ckpt_path, config=local_diffusers_config, safety_checker=None, controlnet=controlnet, local_files_only=True, ) super()._compare_component_configs(pipe, pipe_single_file) def test_single_file_setting_pipeline_dtype_to_fp16(self): controlnet = ControlNetModel.from_pretrained( "diffusers/controlnet-depth-sdxl-1.0", torch_dtype=torch.float16, variant="fp16" ) single_file_pipe = self.pipeline_class.from_single_file( self.ckpt_path, controlnet=controlnet, safety_checker=None, torch_dtype=torch.float16 ) super().test_single_file_setting_pipeline_dtype_to_fp16(single_file_pipe)