import gradio as gr import numpy as np import torch from diffusers.utils import load_image, make_image_grid from diffusers import ( StableDiffusionPipeline, StableDiffusionControlNetPipeline, ControlNetModel ) from peft import PeftModel, LoraConfig from controlnet_aux import HEDdetector from PIL import Image import cv2 as cv import os MAX_SEED = np.iinfo(np.int32).max MAX_IMAGE_SIZE = 1024 IP_ADAPTER = 'h94/IP-Adapter' IP_ADAPTER_WEIGHT_NAME = "ip-adapter-plus_sd15.bin" device = torch.device("cuda" if torch.cuda.is_available() else "cpu") model_id_default = "CompVis/stable-diffusion-v1-4" torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32 hed = None dict_controlnet = { "edge_detection": "lllyasviel/sd-controlnet-canny", # "pose_estimation": "lllyasviel/sd-controlnet-openpose", # "depth_map": "lllyasviel/sd-controlnet-depth", "scribble": "lllyasviel/sd-controlnet-scribble", # "MLSD": "lllyasviel/sd-controlnet-mlsd" } controlnet = ControlNetModel.from_pretrained( dict_controlnet["edge_detection"], cache_dir="./models_cache", torch_dtype=torch_dtype, ) def get_lora_sd_pipeline( ckpt_dir='./lora_logos', base_model_name_or_path=None, dtype=torch.float16, adapter_name="default", controlnet=None ): unet_sub_dir = os.path.join(ckpt_dir, "unet") text_encoder_sub_dir = os.path.join(ckpt_dir, "text_encoder") if os.path.exists(text_encoder_sub_dir) and base_model_name_or_path is None: config = LoraConfig.from_pretrained(text_encoder_sub_dir) base_model_name_or_path = config.base_model_name_or_path if base_model_name_or_path is None: raise ValueError("Please specify the base model name or path") pipe = StableDiffusionControlNetPipeline.from_pretrained( base_model_name_or_path, torch_dtype=dtype, controlnet=controlnet, ) before_params = pipe.unet.parameters() pipe.unet = PeftModel.from_pretrained(pipe.unet, unet_sub_dir, adapter_name=adapter_name) pipe.unet.set_adapter(adapter_name) after_params = pipe.unet.parameters() print("Parameters changed:", any(torch.any(b != a) for b, a in zip(before_params, after_params))) if os.path.exists(text_encoder_sub_dir): pipe.text_encoder = PeftModel.from_pretrained(pipe.text_encoder, text_encoder_sub_dir, adapter_name=adapter_name) if dtype in (torch.float16, torch.bfloat16): pipe.unet.half() pipe.text_encoder.half() return pipe def process_prompt(prompt, tokenizer, text_encoder, max_length=77): tokens = tokenizer(prompt, truncation=False, return_tensors="pt")["input_ids"] chunks = [tokens[:, i:i + max_length] for i in range(0, tokens.shape[1], max_length)] with torch.no_grad(): embeds = [text_encoder(chunk.to(text_encoder.device))[0] for chunk in chunks] return torch.cat(embeds, dim=1) def align_embeddings(prompt_embeds, negative_prompt_embeds): max_length = max(prompt_embeds.shape[1], negative_prompt_embeds.shape[1]) return torch.nn.functional.pad(prompt_embeds, (0, 0, 0, max_length - prompt_embeds.shape[1])), \ torch.nn.functional.pad(negative_prompt_embeds, (0, 0, 0, max_length - negative_prompt_embeds.shape[1])) def map_edge_detection(image_path: str) -> Image: source_img = load_image(image_path).convert('RGB') edges = cv.Canny(np.array(source_img), 80, 160) edges = np.repeat(edges[:, :, None], 3, axis=2) final_image = Image.fromarray(edges) return final_image def map_scribble(image_path: str) -> Image: global hed if not hed: hed = HEDdetector.from_pretrained('lllyasviel/Annotators') image = load_image(image_path).convert('RGB') scribble_image = hed(image) image_np = np.array(scribble_image) image_np = cv.medianBlur(image_np, 3) image = cv.convertScaleAbs(image_np, alpha=1.5, beta=0) final_image = Image.fromarray(image) return final_image pipe = get_lora_sd_pipeline( ckpt_dir='./lora_logos', base_model_name_or_path=model_id_default, dtype=torch_dtype, controlnet=controlnet ).to(device) def infer( prompt, negative_prompt, width=512, height=512, num_inference_steps=20, model_id='CompVis/stable-diffusion-v1-4', seed=42, guidance_scale=7.0, lora_scale=0.5, cn_enable=False, cn_strength=0.0, cn_mode='edge_detection', cn_image=None, ip_enable=False, ip_scale=0.5, ip_image=None, progress=gr.Progress(track_tqdm=True) ): generator = torch.Generator(device).manual_seed(seed) global pipe global controlnet controlnet_changed = False if cn_enable: if dict_controlnet[cn_mode] != pipe.controlnet._name_or_path: controlnet = ControlNetModel.from_pretrained( dict_controlnet[cn_mode], cache_dir="./models_cache", torch_dtype=torch_dtype ) controlnet_changed = True else: cn_strength = 0.0 # отключаем контролнет принудительно if model_id != pipe._name_or_path: pipe = StableDiffusionControlNetPipeline.from_pretrained( model_id, torch_dtype=torch_dtype, controlnet=controlnet, controlnet_conditioning_scale=cn_strength, ).to(device) elif (model_id == pipe._name_or_path) and controlnet_changed: pipe = StableDiffusionControlNetPipeline.from_pretrained( model_id, torch_dtype=torch_dtype, controlnet=controlnet, controlnet_conditioning_scale=cn_strength, ).to(device) print(f"LoRA adapter loaded: {pipe.unet.active_adapters}") print(f"LoRA scale applied: {lora_scale}") pipe.fuse_lora(lora_scale=lora_scale) elif (model_id == pipe._name_or_path) and not controlnet_changed: print(f"LoRA adapter loaded: {pipe.unet.active_adapters}") print(f"LoRA scale applied: {lora_scale}") pipe.fuse_lora(lora_scale=lora_scale) prompt_embeds = process_prompt(prompt, pipe.tokenizer, pipe.text_encoder) negative_prompt_embeds = process_prompt(negative_prompt, pipe.tokenizer, pipe.text_encoder) prompt_embeds, negative_prompt_embeds = align_embeddings(prompt_embeds, negative_prompt_embeds) params = { 'prompt_embeds': prompt_embeds, 'negative_prompt_embeds': negative_prompt_embeds, 'guidance_scale': guidance_scale, 'num_inference_steps': num_inference_steps, 'width': width, 'height': height, 'generator': generator, } if cn_enable: params['controlnet_conditioning_scale'] = cn_strength if cn_mode == 'edge_detection': control_image = map_edge_detection(cn_image) print(type(control_image)) elif cn_mode == 'scribble': control_image = map_scribble(cn_image) params['control_image'] = control_image if ip_enable: pipe.load_ip_adapter( IP_ADAPTER, subfolder="models", weight_name=IP_ADAPTER_WEIGHT_NAME, ) params['ip_adapter_image'] = load_image(ip_image).convert('RGB') pipe.ip_scale(0.6) return pipe(**params).images[0] css = """ #col-container { margin: 0 auto; max-width: 640px; } """ with gr.Blocks(css=css) as demo: with gr.Column(elem_id="col-container"): gr.Markdown(" # DEMO Text-to-Image") with gr.Row(): model_id = gr.Textbox( label="Model ID", max_lines=1, placeholder="Enter model id like 'CompVis/stable-diffusion-v1-4'", value=model_id_default ) prompt = gr.Textbox( label="Prompt", max_lines=1, placeholder="Enter your prompt", ) negative_prompt = gr.Textbox( label="Negative prompt", max_lines=1, placeholder="Enter a negative prompt", ) with gr.Row(): seed = gr.Number( label="Seed", minimum=0, maximum=MAX_SEED, step=1, value=42, ) with gr.Row(): guidance_scale = gr.Slider( label="Guidance scale", minimum=0.0, maximum=10.0, step=0.1, value=7.0, ) with gr.Row(): lora_scale = gr.Slider( label="LoRA scale", minimum=0.0, maximum=1.0, step=0.1, value=0.5, ) with gr.Row(): num_inference_steps = gr.Slider( label="Number of inference steps", minimum=1, maximum=50, step=1, value=20, ) # Секция Control Net cn_enable = gr.Checkbox(label="Enable ControlNet") with gr.Column(visible=False) as cn_options: with gr.Row(): cn_strength = gr.Slider(0, 2, value=0.8, step=0.1, label="Control strength", interactive=True) cn_mode = gr.Dropdown( choices=["edge_detection", "scribble"], value="edge_detection", label="Work regime", interactive=True, ) cn_image = gr.Image(type="filepath", label="Control image") cn_enable.change( lambda x: gr.update(visible=x), inputs=cn_enable, outputs=cn_options ) # Секция IP-Adapter ip_enable = gr.Checkbox(label="Enable IP-Adapter") with gr.Column(visible=False) as ip_options: ip_scale = gr.Slider(0, 1, value=0.5, step=0.1, label="IP-adapter scale", interactive=True) ip_image = gr.Image(type="filepath", label="IP-adapter image", interactive=True) ip_enable.change( lambda x: gr.update(visible=x), inputs=ip_enable, outputs=ip_options ) with gr.Accordion("Optional Settings", open=False): with gr.Row(): width = gr.Slider( label="Width", minimum=256, maximum=MAX_IMAGE_SIZE, step=32, value=512, ) with gr.Row(): height = gr.Slider( label="Height", minimum=256, maximum=MAX_IMAGE_SIZE, step=32, value=512, ) run_button = gr.Button("Run", scale=1, variant="primary") result = gr.Image(label="Result", show_label=False) gr.on( triggers=[run_button.click, prompt.submit], fn=infer, inputs=[ prompt, negative_prompt, width, height, num_inference_steps, model_id, seed, guidance_scale, lora_scale, cn_enable, cn_strength, cn_mode, cn_image, ip_enable, ip_scale, ip_image ], outputs=[result], ) if __name__ == "__main__": demo.launch()