import cv2 import torch import random import numpy as np import spaces import PIL from PIL import Image import diffusers from diffusers.utils import load_image from diffusers.models import ControlNetModel from diffusers.pipelines.controlnet.multicontrolnet import MultiControlNetModel from huggingface_hub import hf_hub_download from insightface.app import FaceAnalysis from style_template import styles from pipeline_stable_diffusion_xl_instantid_full import StableDiffusionXLInstantIDPipeline, draw_kps from controlnet_aux import OpenposeDetector from transformers import DPTImageProcessor, DPTForDepthEstimation import gradio as gr def get_depth_map(image): image = feature_extractor(images=image, return_tensors="pt").pixel_values.to("cuda") with torch.no_grad(), torch.autocast("cuda"): depth_map = depth_estimator(image).predicted_depth depth_map = torch.nn.functional.interpolate( depth_map.unsqueeze(1), size=(1024, 1024), mode="bicubic", align_corners=False, ) depth_min = torch.amin(depth_map, dim=[1, 2, 3], keepdim=True) depth_max = torch.amax(depth_map, dim=[1, 2, 3], keepdim=True) depth_map = (depth_map - depth_min) / (depth_max - depth_min) image = torch.cat([depth_map] * 3, dim=1) image = image.permute(0, 2, 3, 1).cpu().numpy()[0] image = Image.fromarray((image * 255.0).clip(0, 255).astype(np.uint8)) return image def get_canny_image(image, t1=100, t2=200): image = cv2.cvtColor(np.array(image), cv2.COLOR_RGB2BGR) edges = cv2.Canny(image, t1, t2) return Image.fromarray(edges, "L") # global variable MAX_SEED = np.iinfo(np.int32).max device = "cuda" if torch.cuda.is_available() else "cpu" dtype = torch.float16 if str(device).__contains__("cuda") else torch.float32 STYLE_NAMES = list(styles.keys()) DEFAULT_STYLE_NAME = "Watercolor" # download checkpoints from huggingface_hub import hf_hub_download hf_hub_download(repo_id="InstantX/InstantID", filename="ControlNetModel/config.json", local_dir="./checkpoints") hf_hub_download( repo_id="InstantX/InstantID", filename="ControlNetModel/diffusion_pytorch_model.safetensors", local_dir="./checkpoints", ) hf_hub_download(repo_id="InstantX/InstantID", filename="ip-adapter.bin", local_dir="./checkpoints") # Load face encoder app = FaceAnalysis( name="antelopev2", root="./", providers=["CPUExecutionProvider"], ) app.prepare(ctx_id=0, det_size=(640, 640)) depth_estimator = DPTForDepthEstimation.from_pretrained("Intel/dpt-hybrid-midas").to(device) feature_extractor = DPTImageProcessor.from_pretrained("Intel/dpt-hybrid-midas") openpose = OpenposeDetector.from_pretrained("lllyasviel/ControlNet") # Path to InstantID models face_adapter = f"./checkpoints/ip-adapter.bin" controlnet_path = f"./checkpoints/ControlNetModel" # Load pipeline face ControlNetModel controlnet_identitynet = ControlNetModel.from_pretrained( controlnet_path, torch_dtype=dtype ) # controlnet-pose/canny/depth controlnet_pose_model = "thibaud/controlnet-openpose-sdxl-1.0" controlnet_canny_model = "diffusers/controlnet-canny-sdxl-1.0" controlnet_depth_model = "diffusers/controlnet-depth-sdxl-1.0-small" controlnet_pose = ControlNetModel.from_pretrained( controlnet_pose_model, torch_dtype=dtype ).to(device) controlnet_canny = ControlNetModel.from_pretrained( controlnet_canny_model, torch_dtype=dtype ).to(device) controlnet_depth = ControlNetModel.from_pretrained( controlnet_depth_model, torch_dtype=dtype ).to(device) controlnet_map = { "pose": controlnet_pose, "canny": controlnet_canny, "depth": controlnet_depth, } controlnet_map_fn = { "pose": openpose, "canny": get_canny_image, "depth": get_depth_map, } pretrained_model_name_or_path = "wangqixun/YamerMIX_v8" pipe = StableDiffusionXLInstantIDPipeline.from_pretrained( pretrained_model_name_or_path, controlnet=[controlnet_identitynet], torch_dtype=dtype, safety_checker=None, feature_extractor=None, ).to(device) pipe.scheduler = diffusers.EulerDiscreteScheduler.from_config( pipe.scheduler.config ) pipe.load_ip_adapter_instantid(face_adapter) # load and disable LCM pipe.load_lora_weights("latent-consistency/lcm-lora-sdxl") pipe.disable_lora() def toggle_lcm_ui(value): if value: return ( gr.update(minimum=0, maximum=100, step=1, value=5), gr.update(minimum=0.1, maximum=20.0, step=0.1, value=1.5), ) else: return ( gr.update(minimum=5, maximum=100, step=1, value=30), gr.update(minimum=0.1, maximum=20.0, step=0.1, value=5), ) def randomize_seed_fn(seed: int, randomize_seed: bool) -> int: if randomize_seed: seed = random.randint(0, MAX_SEED) return seed def remove_tips(): return gr.update(visible=False) def get_example(): case = [ [ "./examples/yann-lecun_resize.jpg", None, "a man", "Snow", "(lowres, low quality, worst quality:1.2), (text:1.2), watermark, (frame:1.2), deformed, ugly, deformed eyes, blur, out of focus, blurry, deformed cat, deformed, photo, anthropomorphic cat, monochrome, photo, pet collar, gun, weapon, blue, 3d, drones, drone, buildings in background, green", ], [ "./examples/musk_resize.jpeg", "./examples/poses/pose2.jpg", "a man flying in the sky in Mars", "Mars", "(lowres, low quality, worst quality:1.2), (text:1.2), watermark, (frame:1.2), deformed, ugly, deformed eyes, blur, out of focus, blurry, deformed cat, deformed, photo, anthropomorphic cat, monochrome, photo, pet collar, gun, weapon, blue, 3d, drones, drone, buildings in background, green", ], [ "./examples/sam_resize.png", "./examples/poses/pose4.jpg", "a man doing a silly pose wearing a suite", "Jungle", "(lowres, low quality, worst quality:1.2), (text:1.2), watermark, (frame:1.2), deformed, ugly, deformed eyes, blur, out of focus, blurry, deformed cat, deformed, photo, anthropomorphic cat, monochrome, photo, pet collar, gun, weapon, blue, 3d, drones, drone, buildings in background, gree", ], [ "./examples/schmidhuber_resize.png", "./examples/poses/pose3.jpg", "a man sit on a chair", "Neon", "(lowres, low quality, worst quality:1.2), (text:1.2), watermark, (frame:1.2), deformed, ugly, deformed eyes, blur, out of focus, blurry, deformed cat, deformed, photo, anthropomorphic cat, monochrome, photo, pet collar, gun, weapon, blue, 3d, drones, drone, buildings in background, green", ], [ "./examples/kaifu_resize.png", "./examples/poses/pose.jpg", "a man", "Vibrant Color", "(lowres, low quality, worst quality:1.2), (text:1.2), watermark, (frame:1.2), deformed, ugly, deformed eyes, blur, out of focus, blurry, deformed cat, deformed, photo, anthropomorphic cat, monochrome, photo, pet collar, gun, weapon, blue, 3d, drones, drone, buildings in background, green", ], ] return case def run_for_examples(face_file, pose_file, prompt, style, negative_prompt): return generate_image( face_file, pose_file, prompt, negative_prompt, style, 20, # num_steps 0.8, # identitynet_strength_ratio 0.8, # adapter_strength_ratio 0.4, # pose_strength 0.3, # canny_strength 0.5, # depth_strength ["pose", "canny"], # controlnet_selection 5.0, # guidance_scale 42, # seed "EulerDiscreteScheduler", # scheduler False, # enable_LCM True, # enable_Face_Region ) def convert_from_cv2_to_image(img: np.ndarray) -> Image: return Image.fromarray(cv2.cvtColor(img, cv2.COLOR_BGR2RGB)) def convert_from_image_to_cv2(img: Image) -> np.ndarray: return cv2.cvtColor(np.array(img), cv2.COLOR_RGB2BGR) def resize_img( input_image, max_side=1280, min_side=1024, size=None, pad_to_max_side=False, mode=PIL.Image.BILINEAR, base_pixel_number=64, ): w, h = input_image.size if size is not None: w_resize_new, h_resize_new = size else: ratio = min_side / min(h, w) w, h = round(ratio * w), round(ratio * h) ratio = max_side / max(h, w) input_image = input_image.resize([round(ratio * w), round(ratio * h)], mode) w_resize_new = (round(ratio * w) // base_pixel_number) * base_pixel_number h_resize_new = (round(ratio * h) // base_pixel_number) * base_pixel_number input_image = input_image.resize([w_resize_new, h_resize_new], mode) if pad_to_max_side: res = np.ones([max_side, max_side, 3], dtype=np.uint8) * 255 offset_x = (max_side - w_resize_new) // 2 offset_y = (max_side - h_resize_new) // 2 res[ offset_y : offset_y + h_resize_new, offset_x : offset_x + w_resize_new ] = np.array(input_image) input_image = Image.fromarray(res) return input_image def apply_style( style_name: str, positive: str, negative: str = "" ) -> tuple[str, str]: p, n = styles.get(style_name, styles[DEFAULT_STYLE_NAME]) return p.replace("{prompt}", positive), n + " " + negative @spaces.GPU def generate_image( face_image_path, pose_image_path, prompt, negative_prompt, style_name, num_steps, identitynet_strength_ratio, adapter_strength_ratio, pose_strength, canny_strength, depth_strength, controlnet_selection, guidance_scale, seed, scheduler, enable_LCM, enhance_face_region, progress=gr.Progress(track_tqdm=True), ): if enable_LCM: pipe.scheduler = diffusers.LCMScheduler.from_config(pipe.scheduler.config) pipe.enable_lora() else: pipe.disable_lora() scheduler_class_name = scheduler.split("-")[0] add_kwargs = {} if len(scheduler.split("-")) > 1: add_kwargs["use_karras_sigmas"] = True if len(scheduler.split("-")) > 2: add_kwargs["algorithm_type"] = "sde-dpmsolver++" scheduler = getattr(diffusers, scheduler_class_name) pipe.scheduler = scheduler.from_config(pipe.scheduler.config, **add_kwargs) if face_image_path is None: raise gr.Error( f"Cannot find any input face image! Please upload the face image" ) if prompt is None: prompt = "a person" # apply the style template prompt, negative_prompt = apply_style(style_name, prompt, negative_prompt) face_image = load_image(face_image_path) face_image = resize_img(face_image, max_side=1024) face_image_cv2 = convert_from_image_to_cv2(face_image) height, width, _ = face_image_cv2.shape # Extract face features face_info = app.get(face_image_cv2) if len(face_info) == 0: raise gr.Error( f"Unable to detect a face in the image. Please upload a different photo with a clear face." ) face_info = sorted( face_info, key=lambda x: (x["bbox"][2] - x["bbox"][0]) * x["bbox"][3] - x["bbox"][1], )[ -1 ] # only use the maximum face face_emb = face_info["embedding"] face_kps = draw_kps(convert_from_cv2_to_image(face_image_cv2), face_info["kps"]) img_controlnet = face_image if pose_image_path is not None: pose_image = load_image(pose_image_path) pose_image = resize_img(pose_image, max_side=1024) img_controlnet = pose_image pose_image_cv2 = convert_from_image_to_cv2(pose_image) face_info = app.get(pose_image_cv2) if len(face_info) == 0: raise gr.Error( f"Cannot find any face in the reference image! Please upload another person image" ) face_info = face_info[-1] face_kps = draw_kps(pose_image, face_info["kps"]) width, height = face_kps.size if enhance_face_region: control_mask = np.zeros([height, width, 3]) x1, y1, x2, y2 = face_info["bbox"] x1, y1, x2, y2 = int(x1), int(y1), int(x2), int(y2) control_mask[y1:y2, x1:x2] = 255 control_mask = Image.fromarray(control_mask.astype(np.uint8)) else: control_mask = None if len(controlnet_selection) > 0: controlnet_scales = { "pose": pose_strength, "canny": canny_strength, "depth": depth_strength, } pipe.controlnet = MultiControlNetModel( [controlnet_identitynet] + [controlnet_map[s] for s in controlnet_selection] ) control_scales = [float(identitynet_strength_ratio)] + [ controlnet_scales[s] for s in controlnet_selection ] control_images = [face_kps] + [ controlnet_map_fn[s](img_controlnet).resize((width, height)) for s in controlnet_selection ] else: pipe.controlnet = controlnet_identitynet control_scales = float(identitynet_strength_ratio) control_images = face_kps generator = torch.Generator(device=device).manual_seed(seed) print("Start inference...") print(f"[Debug] Prompt: {prompt}, \n[Debug] Neg Prompt: {negative_prompt}") pipe.set_ip_adapter_scale(adapter_strength_ratio) images = pipe( prompt=prompt, negative_prompt=negative_prompt, image_embeds=face_emb, image=control_images, control_mask=control_mask, controlnet_conditioning_scale=control_scales, num_inference_steps=num_steps, guidance_scale=guidance_scale, height=height, width=width, generator=generator, ).images return images[0], gr.update(visible=True) # Description title = r"""

InstantID: Zero-shot Identity-Preserving Generation in Seconds

""" description = r""" Official 🤗 Gradio demo for InstantID: Zero-shot Identity-Preserving Generation in Seconds.
How to use:
1. Upload an image with a face. For images with multiple faces, we will only detect the largest face. Ensure the face is not too small and is clearly visible without significant obstructions or blurring. 2. (Optional) You can upload another image as a reference for the face pose. If you don't, we will use the first detected face image to extract facial landmarks. If you use a cropped face at step 1, it is recommended to upload it to define a new face pose. 3. (Optional) You can select multiple ControlNet models to control the generation process. The default is to use the IdentityNet only. The ControlNet models include pose skeleton, canny, and depth. You can adjust the strength of each ControlNet model to control the generation process. 4. Enter a text prompt, as done in normal text-to-image models. 5. Click the Submit button to begin customization. 6. Share your customized photo with your friends and enjoy! 😊""" article = r""" --- 📝 **Citation**
If our work is helpful for your research or applications, please cite us via: ```bibtex @article{wang2024instantid, title={InstantID: Zero-shot Identity-Preserving Generation in Seconds}, author={Wang, Qixun and Bai, Xu and Wang, Haofan and Qin, Zekui and Chen, Anthony}, journal={arXiv preprint arXiv:2401.07519}, year={2024} } ``` 📧 **Contact**
If you have any questions, please feel free to open an issue or directly reach us out at haofanwang.ai@gmail.com. """ tips = r""" ### Usage tips of InstantID 1. If you're not satisfied with the similarity, try increasing the weight of "IdentityNet Strength" and "Adapter Strength." 2. If you feel that the saturation is too high, first decrease the Adapter strength. If it remains too high, then decrease the IdentityNet strength. 3. If you find that text control is not as expected, decrease Adapter strength. 4. If you find that realistic style is not good enough, go for our Github repo and use a more realistic base model. """ css = """ .gradio-container {width: 85% !important} """ with gr.Blocks(css=css) as demo: # description gr.Markdown(title) gr.Markdown(description) with gr.Row(): with gr.Column(): with gr.Row(equal_height=True): # upload face image face_file = gr.Image( label="Upload a photo of your face", type="filepath" ) # optional: upload a reference pose image pose_file = gr.Image( label="Upload a reference pose image (Optional)", type="filepath", ) # prompt prompt = gr.Textbox( label="Prompt", info="Give simple prompt is enough to achieve good face fidelity", placeholder="A photo of a person", value="", ) submit = gr.Button("Submit", variant="primary") enable_LCM = gr.Checkbox( label="Enable Fast Inference with LCM", value=enable_lcm_arg, info="LCM speeds up the inference step, the trade-off is the quality of the generated image. It performs better with portrait face images rather than distant faces", ) style = gr.Dropdown( label="Style template", choices=STYLE_NAMES, value=DEFAULT_STYLE_NAME, ) # strength identitynet_strength_ratio = gr.Slider( label="IdentityNet strength (for fidelity)", minimum=0, maximum=1.5, step=0.05, value=0.80, ) adapter_strength_ratio = gr.Slider( label="Image adapter strength (for detail)", minimum=0, maximum=1.5, step=0.05, value=0.80, ) with gr.Accordion("Controlnet"): controlnet_selection = gr.CheckboxGroup( ["pose", "canny", "depth"], label="Controlnet", value=["pose"], info="Use pose for skeleton inference, canny for edge detection, and depth for depth map estimation. You can try all three to control the generation process" ) pose_strength = gr.Slider( label="Pose strength", minimum=0, maximum=1.5, step=0.05, value=0.40, ) canny_strength = gr.Slider( label="Canny strength", minimum=0, maximum=1.5, step=0.05, value=0.40, ) depth_strength = gr.Slider( label="Depth strength", minimum=0, maximum=1.5, step=0.05, value=0.40, ) with gr.Accordion(open=False, label="Advanced Options"): negative_prompt = gr.Textbox( label="Negative Prompt", placeholder="low quality", value="(lowres, low quality, worst quality:1.2), (text:1.2), watermark, (frame:1.2), deformed, ugly, deformed eyes, blur, out of focus, blurry, deformed cat, deformed, photo, anthropomorphic cat, monochrome, pet collar, gun, weapon, blue, 3d, drones, drone, buildings in background, green", ) num_steps = gr.Slider( label="Number of sample steps", minimum=1, maximum=100, step=1, value=5 if enable_lcm_arg else 30, ) guidance_scale = gr.Slider( label="Guidance scale", minimum=0.1, maximum=20.0, step=0.1, value=0.0 if enable_lcm_arg else 5.0, ) seed = gr.Slider( label="Seed", minimum=0, maximum=MAX_SEED, step=1, value=42, ) schedulers = [ "DEISMultistepScheduler", "HeunDiscreteScheduler", "EulerDiscreteScheduler", "DPMSolverMultistepScheduler", "DPMSolverMultistepScheduler-Karras", "DPMSolverMultistepScheduler-Karras-SDE", ] scheduler = gr.Dropdown( label="Schedulers", choices=schedulers, value="EulerDiscreteScheduler", ) randomize_seed = gr.Checkbox(label="Randomize seed", value=True) enhance_face_region = gr.Checkbox(label="Enhance non-face region", value=True) with gr.Column(scale=1): gallery = gr.Image(label="Generated Images") usage_tips = gr.Markdown( label="InstantID Usage Tips", value=tips, visible=False ) submit.click( fn=remove_tips, outputs=usage_tips, ).then( fn=randomize_seed_fn, inputs=[seed, randomize_seed], outputs=seed, queue=False, api_name=False, ).then( fn=generate_image, inputs=[ face_file, pose_file, prompt, negative_prompt, style, num_steps, identitynet_strength_ratio, adapter_strength_ratio, pose_strength, canny_strength, depth_strength, controlnet_selection, guidance_scale, seed, scheduler, enable_LCM, enhance_face_region, ], outputs=[gallery, usage_tips], ) enable_LCM.input( fn=toggle_lcm_ui, inputs=[enable_LCM], outputs=[num_steps, guidance_scale], queue=False, ) gr.Examples( examples=get_example(), inputs=[face_file, pose_file, prompt, style, negative_prompt], fn=run_for_examples, outputs=[gallery, usage_tips], cache_examples=True, ) gr.Markdown(article) demo.queue(api_open=False) demo.launch()