import spaces import random import torch import cv2 import insightface import gradio as gr import numpy as np import os from huggingface_hub import snapshot_download from transformers import CLIPVisionModelWithProjection,CLIPImageProcessor from kolors.pipelines.pipeline_stable_diffusion_xl_chatglm_256_ipadapter_FaceID import StableDiffusionXLPipeline from kolors.models.modeling_chatglm import ChatGLMModel from kolors.models.tokenization_chatglm import ChatGLMTokenizer from diffusers import AutoencoderKL from kolors.models.unet_2d_condition import UNet2DConditionModel from diffusers import EulerDiscreteScheduler from PIL import Image from insightface.app import FaceAnalysis from insightface.data import get_image as ins_get_image device = "cuda" ckpt_dir = snapshot_download(repo_id="Kwai-Kolors/Kolors") ckpt_dir_faceid = snapshot_download(repo_id="Kwai-Kolors/Kolors-IP-Adapter-FaceID-Plus") text_encoder = ChatGLMModel.from_pretrained(f'{ckpt_dir}/text_encoder', torch_dtype=torch.float16).half().to(device) tokenizer = ChatGLMTokenizer.from_pretrained(f'{ckpt_dir}/text_encoder') vae = AutoencoderKL.from_pretrained(f"{ckpt_dir}/vae", revision=None).half().to(device) scheduler = EulerDiscreteScheduler.from_pretrained(f"{ckpt_dir}/scheduler") unet = UNet2DConditionModel.from_pretrained(f"{ckpt_dir}/unet", revision=None).half().to(device) clip_image_encoder = CLIPVisionModelWithProjection.from_pretrained(f'{ckpt_dir_faceid}/clip-vit-large-patch14-336', ignore_mismatched_sizes=True) clip_image_encoder.to(device) clip_image_processor = CLIPImageProcessor(size = 336, crop_size = 336) pipe = StableDiffusionXLPipeline( vae = vae, text_encoder = text_encoder, tokenizer = tokenizer, unet = unet, scheduler = scheduler, face_clip_encoder = clip_image_encoder, face_clip_processor = clip_image_processor, force_zeros_for_empty_prompt = False, ) class FaceInfoGenerator(): def __init__(self, root_dir = "./.insightface/"): self.app = FaceAnalysis(name = 'antelopev2', root = root_dir, providers=['CUDAExecutionProvider', 'CPUExecutionProvider']) self.app.prepare(ctx_id = 0, det_size = (640, 640)) def get_faceinfo_one_img(self, face_image): face_info = self.app.get(cv2.cvtColor(np.array(face_image), cv2.COLOR_RGB2BGR)) if len(face_info) == 0: face_info = None else: 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 return face_info def face_bbox_to_square(bbox): ## l, t, r, b to square l, t, r, b l,t,r,b = bbox cent_x = (l + r) / 2 cent_y = (t + b) / 2 w, h = r - l, b - t r = max(w, h) / 2 l0 = cent_x - r r0 = cent_x + r t0 = cent_y - r b0 = cent_y + r return [l0, t0, r0, b0] MAX_SEED = np.iinfo(np.int32).max MAX_IMAGE_SIZE = 1024 face_info_generator = FaceInfoGenerator() @spaces.GPU def infer(prompt, image = None, negative_prompt = "nsfw，脸部阴影，低分辨率，jpeg伪影、模糊、糟糕，黑脸，霓虹灯", seed = 66, randomize_seed = False, guidance_scale = 5.0, num_inference_steps = 50 ): if randomize_seed: seed = random.randint(0, MAX_SEED) generator = torch.Generator().manual_seed(seed) global pipe pipe = pipe.to(device) pipe.load_ip_adapter_faceid_plus(f'{ckpt_dir_faceid}/ipa-faceid-plus.bin', device = device) scale = 0.8 pipe.set_face_fidelity_scale(scale) face_info = face_info_generator.get_faceinfo_one_img(image) face_bbox_square = face_bbox_to_square(face_info["bbox"]) crop_image = image.crop(face_bbox_square) crop_image = crop_image.resize((336, 336)) crop_image = [crop_image] face_embeds = torch.from_numpy(np.array([face_info["embedding"]])) face_embeds = face_embeds.to(device, dtype = torch.float16) image = pipe( prompt = prompt, negative_prompt = negative_prompt, height = 1024, width = 1024, num_inference_steps= num_inference_steps, guidance_scale = guidance_scale, num_images_per_prompt = 1, generator = generator, face_crop_image = crop_image, face_insightface_embeds = face_embeds ).images[0] return image, seed examples = [ ["穿着晚礼服，在星光下的晚宴场景中，烛光闪闪，整个场景洋溢着浪漫而奢华的氛围", "image/image1.png"], ["西部牛仔，牛仔帽，荒野大镖客，背景是西部小镇，仙人掌，,日落余晖, 暖色调, 使用XT4胶片拍摄, 噪点, 晕影, 柯达胶卷，复古", "image/image2.png"] ] css=""" #col-left { margin: 0 auto; max-width: 600px; } #col-right { margin: 0 auto; max-width: 750px; } #button { color: blue; } """ def load_description(fp): with open(fp, 'r', encoding='utf-8') as f: content = f.read() return content with gr.Blocks(css=css) as Kolors: gr.HTML(load_description("assets/title.md")) with gr.Row(): with gr.Column(elem_id="col-left"): with gr.Row(): prompt = gr.Textbox( label="Prompt", placeholder="Enter your prompt", lines=2 ) with gr.Row(): image = gr.Image(label="Image", type="pil") with gr.Accordion("Advanced Settings", open=False): negative_prompt = gr.Textbox( label="Negative prompt", placeholder="Enter a negative prompt", visible=True, ) seed = gr.Slider( label="Seed", minimum=0, maximum=MAX_SEED, step=1, value=0, ) randomize_seed = gr.Checkbox(label="Randomize seed", value=True) with gr.Row(): guidance_scale = gr.Slider( label="Guidance scale", minimum=0.0, maximum=10.0, step=0.1, value=5.0, ) num_inference_steps = gr.Slider( label="Number of inference steps", minimum=10, maximum=50, step=1, value=25, ) with gr.Row(): button = gr.Button("Run", elem_id="button") with gr.Column(elem_id="col-right"): result = gr.Image(label="Result", show_label=False) seed_used = gr.Number(label="Seed Used") with gr.Row(): gr.Examples( fn = infer, examples = examples, inputs = [prompt, image], outputs = [result, seed_used], ) button.click( fn = infer, inputs = [prompt, image, negative_prompt, seed, randomize_seed, guidance_scale, num_inference_steps], outputs = [result, seed_used] ) Kolors.queue().launch(debug=True)