import spaces import random import torch import cv2 import gradio as gr import numpy as np from huggingface_hub import snapshot_download from transformers import CLIPVisionModelWithProjection, CLIPImageProcessor from diffusers.utils import load_image from kolors.pipelines.pipeline_controlnet_xl_kolors_img2img import StableDiffusionXLControlNetImg2ImgPipeline from kolors.models.modeling_chatglm import ChatGLMModel from kolors.models.tokenization_chatglm import ChatGLMTokenizer from kolors.models.controlnet import ControlNetModel from diffusers import AutoencoderKL from kolors.models.unet_2d_condition import UNet2DConditionModel from diffusers import EulerDiscreteScheduler from PIL import Image from annotator.midas import MidasDetector from annotator.dwpose import DWposeDetector from annotator.util import resize_image, HWC3 from transformers import pipeline device = "cuda" ckpt_dir = snapshot_download(repo_id="Kwai-Kolors/Kolors") ckpt_dir_depth = snapshot_download(repo_id="Kwai-Kolors/Kolors-ControlNet-Depth") ckpt_dir_canny = snapshot_download(repo_id="Kwai-Kolors/Kolors-ControlNet-Canny") ckpt_dir_pose = snapshot_download(repo_id="Kwai-Kolors/Kolors-ControlNet-Pose") 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) controlnet_depth = ControlNetModel.from_pretrained(f"{ckpt_dir_depth}", revision=None).half().to(device) controlnet_canny = ControlNetModel.from_pretrained(f"{ckpt_dir_canny}", revision=None).half().to(device) controlnet_pose = ControlNetModel.from_pretrained(f"{ckpt_dir_pose}", revision=None).half().to(device) pipe_depth = StableDiffusionXLControlNetImg2ImgPipeline( vae=vae, controlnet=controlnet_depth, text_encoder=text_encoder, tokenizer=tokenizer, unet=unet, scheduler=scheduler, force_zeros_for_empty_prompt=False ) pipe_canny = StableDiffusionXLControlNetImg2ImgPipeline( vae=vae, controlnet=controlnet_canny, text_encoder=text_encoder, tokenizer=tokenizer, unet=unet, scheduler=scheduler, force_zeros_for_empty_prompt=False ) pipe_pose = StableDiffusionXLControlNetImg2ImgPipeline( vae=vae, controlnet=controlnet_pose, text_encoder=text_encoder, tokenizer=tokenizer, unet=unet, scheduler=scheduler, force_zeros_for_empty_prompt=False ) # 번역 모델 초기화 translator = pipeline("translation", model="Helsinki-NLP/opus-mt-ko-en") # prompt를 처리하는 함수 추가 def process_prompt(prompt): if any('\u3131' <= char <= '\u3163' or '\uac00' <= char <= '\ud7a3' for char in prompt): translated = translator(prompt)[0]['translation_text'] return prompt, translated return prompt, prompt @spaces.GPU def process_canny_condition(image, canny_threods=[100,200]): np_image = image.copy() np_image = cv2.Canny(np_image, canny_threods[0], canny_threods[1]) np_image = np_image[:, :, None] np_image = np.concatenate([np_image, np_image, np_image], axis=2) np_image = HWC3(np_image) return Image.fromarray(np_image) model_midas = MidasDetector() @spaces.GPU def process_depth_condition_midas(img, res = 1024): h,w,_ = img.shape img = resize_image(HWC3(img), res) result = HWC3(model_midas(img)) result = cv2.resize(result, (w,h)) return Image.fromarray(result) model_dwpose = DWposeDetector() @spaces.GPU def process_dwpose_condition(image, res=1024): h,w,_ = image.shape img = resize_image(HWC3(image), res) out_res, out_img = model_dwpose(image) result = HWC3(out_img) result = cv2.resize(result, (w,h)) return Image.fromarray(result) MAX_SEED = np.iinfo(np.int32).max MAX_IMAGE_SIZE = 1024 @spaces.GPU def infer_depth(prompt, image = None, negative_prompt = "NSFW, facial shadow, low resolution, JPEG artifacts, blurry, poor quality, blackface, neon lights.", seed = 397886929, randomize_seed = False, guidance_scale = 6.0, num_inference_steps = 50, controlnet_conditioning_scale = 0.7, control_guidance_end = 0.9, strength = 1.0 ): original_prompt, english_prompt = process_prompt(prompt) if randomize_seed: seed = random.randint(0, MAX_SEED) generator = torch.Generator().manual_seed(seed) init_image = resize_image(image, MAX_IMAGE_SIZE) pipe = pipe_depth.to("cuda") condi_img = process_depth_condition_midas(np.array(init_image), MAX_IMAGE_SIZE) image = pipe( prompt=english_prompt, image=init_image, controlnet_conditioning_scale=controlnet_conditioning_scale, control_guidance_end=control_guidance_end, strength=strength, control_image=condi_img, negative_prompt=negative_prompt, num_inference_steps=num_inference_steps, guidance_scale=guidance_scale, num_images_per_prompt=1, generator=generator, ).images[0] return [condi_img, image], seed, original_prompt, english_prompt @spaces.GPU def infer_canny(prompt, image = None, negative_prompt = "NSFW, facial shadow, low resolution, JPEG artifacts, blurry, poor quality, blackface, neon lights.", seed = 397886929, randomize_seed = False, guidance_scale = 6.0, num_inference_steps = 50, controlnet_conditioning_scale = 0.7, control_guidance_end = 0.9, strength = 1.0 ): original_prompt, english_prompt = process_prompt(prompt) if randomize_seed: seed = random.randint(0, MAX_SEED) generator = torch.Generator().manual_seed(seed) init_image = resize_image(image, MAX_IMAGE_SIZE) pipe = pipe_canny.to("cuda") condi_img = process_canny_condition(np.array(init_image)) image = pipe( prompt=english_prompt, image=init_image, controlnet_conditioning_scale=controlnet_conditioning_scale, control_guidance_end=control_guidance_end, strength=strength, control_image=condi_img, negative_prompt=negative_prompt, num_inference_steps=num_inference_steps, guidance_scale=guidance_scale, num_images_per_prompt=1, generator=generator, ).images[0] return [condi_img, image], seed, original_prompt, english_prompt @spaces.GPU def infer_pose(prompt, image = None, negative_prompt = "NSFW, facial shadow, low resolution, JPEG artifacts, blurry, poor quality, blackface, neon lights.", seed = 66, randomize_seed = False, guidance_scale = 6.0, num_inference_steps = 50, controlnet_conditioning_scale = 0.7, control_guidance_end = 0.9, strength = 1.0 ): original_prompt, english_prompt = process_prompt(prompt) if randomize_seed: seed = random.randint(0, MAX_SEED) generator = torch.Generator().manual_seed(seed) init_image = resize_image(image, MAX_IMAGE_SIZE) pipe = pipe_pose.to("cuda") condi_img = process_dwpose_condition(np.array(init_image), MAX_IMAGE_SIZE) image = pipe( prompt=english_prompt, image=init_image, controlnet_conditioning_scale=controlnet_conditioning_scale, control_guidance_end=control_guidance_end, strength=strength, control_image=condi_img, negative_prompt=negative_prompt, num_inference_steps=num_inference_steps, guidance_scale=guidance_scale, num_images_per_prompt=1, generator=generator, ).images[0] return [condi_img, image], seed, original_prompt, english_prompt canny_examples = [ ["아름다운 소녀, 고품질, 매우 선명, 생생한 색상, 초고해상도, 최상의 품질, 8k, 고화질, 4K", "image/woman_1.png"], ["파노라마, 컵 안에 앉아있는 귀여운 흰 강아지, 카메라를 바라보는, 애니메이션 스타일, 3D 렌더링, 옥테인 렌더", "image/dog.png"] ] depth_examples = [ ["신카이 마코토 스타일, 풍부한 색감, 초록 셔츠를 입은 여성이 들판에 서 있는, 아름다운 풍경, 맑고 밝은, 얼룩진 빛과 그림자, 최고의 품질, 초세밀, 8K 화질", "image/woman_2.png"], ["화려한 색상의 작은 새, 고품질, 매우 선명, 생생한 색상, 초고해상도, 최상의 품질, 8k, 고화질, 4K", "image/bird.png"] ] pose_examples = [ ["보라색 퍼프 슬리브 드레스를 입고 왕관과 흰색 레이스 장갑을 낀 소녀가 양 손으로 얼굴을 감싸고 있는, 고품질, 매우 선명, 생생한 색상, 초고해상도, 최상의 품질, 8k, 고화질, 4K", "image/woman_3.png"], ["검은색 스포츠 재킷과 흰색 이너를 입고 목걸이를 한 여성이 거리에 서 있는, 배경은 빨간 건물과 녹색 나무, 고품질, 매우 선명, 생생한 색상, 초고해상도, 최상의 품질, 8k, 고화질, 4K", "image/woman_4.png"] ] css = """ footer { visibility: hidden; } """ def load_description(fp): with open(fp, 'r', encoding='utf-8') as f: content = f.read() return content with gr.Blocks(theme="Nymbo/Nymbo_Theme", css=css) as Kolors: with gr.Row(): with gr.Column(elem_id="col-left"): with gr.Row(): prompt = gr.Textbox( label="프롬프트", placeholder="프롬프트를 입력하세요 (한글 또는 영어)", lines=2 ) with gr.Row(): image = gr.Image(label="이미지", type="pil") with gr.Accordion("고급 설정", open=False): negative_prompt = gr.Textbox( label="네거티브 프롬프트", placeholder="네거티브 프롬프트를 입력하세요", visible=True, value="nsfw, 얼굴 그림자, 저해상도, jpeg 아티팩트, 흐릿함, 열악함, 검은 얼굴, 네온 조명" ) seed = gr.Slider( label="시드", minimum=0, maximum=MAX_SEED, step=1, value=0, ) randomize_seed = gr.Checkbox(label="시드 무작위화", value=True) with gr.Row(): guidance_scale = gr.Slider( label="가이던스 스케일", minimum=0.0, maximum=10.0, step=0.1, value=6.0, ) num_inference_steps = gr.Slider( label="추론 단계 수", minimum=10, maximum=50, step=1, value=30, ) with gr.Row(): controlnet_conditioning_scale = gr.Slider( label="컨트롤넷 컨디셔닝 스케일", minimum=0.0, maximum=1.0, step=0.1, value=0.7, ) control_guidance_end = gr.Slider( label="컨트롤 가이던스 종료", minimum=0.0, maximum=1.0, step=0.1, value=0.9, ) with gr.Row(): strength = gr.Slider( label="강도", minimum=0.0, maximum=1.0, step=0.1, value=1.0, ) with gr.Row(): canny_button = gr.Button("캐니", elem_id="button") depth_button = gr.Button("깊이", elem_id="button") pose_button = gr.Button("포즈", elem_id="button") with gr.Column(elem_id="col-right"): result = gr.Gallery(label="결과", show_label=False, columns=2) seed_used = gr.Number(label="사용된 시드") original_prompt_display = gr.Textbox(label="원본 프롬프트") english_prompt_display = gr.Textbox(label="영어 프롬프트") with gr.Row(): gr.Examples( fn=infer_canny, examples=canny_examples, inputs=[prompt, image], outputs=[result, seed_used, original_prompt_display, english_prompt_display], label="Canny" ) with gr.Row(): gr.Examples( fn=infer_depth, examples=depth_examples, inputs=[prompt, image], outputs=[result, seed_used, original_prompt_display, english_prompt_display], label="Depth" ) with gr.Row(): gr.Examples( fn=infer_pose, examples=pose_examples, inputs=[prompt, image], outputs=[result, seed_used, original_prompt_display, english_prompt_display], label="Pose" ) canny_button.click( fn=infer_canny, inputs=[prompt, image, negative_prompt, seed, randomize_seed, guidance_scale, num_inference_steps, controlnet_conditioning_scale, control_guidance_end, strength], outputs=[result, seed_used, original_prompt_display, english_prompt_display] ) depth_button.click( fn=infer_depth, inputs=[prompt, image, negative_prompt, seed, randomize_seed, guidance_scale, num_inference_steps, controlnet_conditioning_scale, control_guidance_end, strength], outputs=[result, seed_used, original_prompt_display, english_prompt_display] ) pose_button.click( fn=infer_pose, inputs=[prompt, image, negative_prompt, seed, randomize_seed, guidance_scale, num_inference_steps, controlnet_conditioning_scale, control_guidance_end, strength], outputs=[result, seed_used, original_prompt_display, english_prompt_display] ) Kolors.queue().launch(debug=True)