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 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_ipa = snapshot_download(repo_id="Kwai-Kolors/Kolors-IP-Adapter-Plus") 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) image_encoder = CLIPVisionModelWithProjection.from_pretrained(f'{ckpt_dir_ipa}/image_encoder', ignore_mismatched_sizes=True).to(dtype=torch.float16, device=device) ip_img_size = 336 clip_image_processor = CLIPImageProcessor(size=ip_img_size, crop_size=ip_img_size ) pipe_depth = StableDiffusionXLControlNetImg2ImgPipeline( vae=vae, controlnet = controlnet_depth, text_encoder=text_encoder, tokenizer=tokenizer, unet=unet, scheduler=scheduler, image_encoder=image_encoder, feature_extractor=clip_image_processor, force_zeros_for_empty_prompt=False ) pipe_canny = StableDiffusionXLControlNetImg2ImgPipeline( vae=vae, controlnet = controlnet_canny, text_encoder=text_encoder, tokenizer=tokenizer, unet=unet, scheduler=scheduler, image_encoder=image_encoder, feature_extractor=clip_image_processor, force_zeros_for_empty_prompt=False ) pipe_pose = StableDiffusionXLControlNetImg2ImgPipeline( vae=vae, controlnet = controlnet_pose, text_encoder=text_encoder, tokenizer=tokenizer, unet=unet, scheduler=scheduler, image_encoder=image_encoder, feature_extractor=clip_image_processor, force_zeros_for_empty_prompt=False ) for pipe in [pipe_depth]: if hasattr(pipe.unet, 'encoder_hid_proj'): pipe.unet.text_encoder_hid_proj = pipe.unet.encoder_hid_proj pipe_depth.load_ip_adapter(f'{ckpt_dir_ipa}' , subfolder="", weight_name=["ip_adapter_plus_general.bin"]) pipe_canny.load_ip_adapter(f'{ckpt_dir_ipa}' , subfolder="", weight_name=["ip_adapter_plus_general.bin"]) pipe_pose.load_ip_adapter(f'{ckpt_dir_ipa}' , subfolder="", weight_name=["ip_adapter_plus_general.bin"]) @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, ipa_img = None, negative_prompt = "nsfw,脸部阴影,低分辨率,糟糕的解剖结构、糟糕的手,缺失手指、质量最差、低质量、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, controlnet_conditioning_scale = 0.5, control_guidance_end = 0.9, strength = 1.0, ip_scale = 0.5, ): 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") pipe.set_ip_adapter_scale([ip_scale]) condi_img = process_depth_condition_midas( np.array(init_image), MAX_IMAGE_SIZE) image = pipe( prompt= prompt , image = init_image, controlnet_conditioning_scale = controlnet_conditioning_scale, control_guidance_end = control_guidance_end, ip_adapter_image=[ipa_img], 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 @spaces.GPU def infer_canny(prompt, image = None, ipa_img = None, negative_prompt = "nsfw,脸部阴影,低分辨率,糟糕的解剖结构、糟糕的手,缺失手指、质量最差、低质量、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, controlnet_conditioning_scale = 0.5, control_guidance_end = 0.9, strength = 1.0, ip_scale = 0.5, ): 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") pipe.set_ip_adapter_scale([ip_scale]) condi_img = process_canny_condition(np.array(init_image)) image = pipe( prompt= prompt , image = init_image, controlnet_conditioning_scale = controlnet_conditioning_scale, control_guidance_end = control_guidance_end, ip_adapter_image=[ipa_img], 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 @spaces.GPU def infer_pose(prompt, image = None, ipa_img = 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, controlnet_conditioning_scale = 0.5, control_guidance_end = 0.9, strength = 1.0, ip_scale = 0.5, ): 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") pipe.set_ip_adapter_scale([ip_scale]) condi_img = process_dwpose_condition(np.array(init_image), MAX_IMAGE_SIZE) image = pipe( prompt= prompt , image = init_image, controlnet_conditioning_scale = controlnet_conditioning_scale, control_guidance_end = control_guidance_end, ip_adapter_image=[ipa_img], 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 canny_examples = [ ["一个红色头发的女孩,唯美风景,清新明亮,斑驳的光影,最好的质量,超细节,8K画质", "image/woman_2.png", "image/2.png"], ] depth_examples = [ ["一个漂亮的女孩,最好的质量,超细节,8K画质", "image/1.png","image/woman_1.png"], ] pose_examples = [ ["一位穿着紫色泡泡袖连衣裙、戴着皇冠和白色蕾丝手套的女孩,超高分辨率,最佳品质,8k画质", "image/woman_3.png","image/woman_4.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") ipa_image = gr.Image(label="IP-Adapter-Image", type="pil") with gr.Accordion("Advanced Settings", open=False): negative_prompt = gr.Textbox( label="Negative prompt", placeholder="Enter a negative prompt", visible=True, value="nsfw,脸部阴影,低分辨率,糟糕的解剖结构、糟糕的手,缺失手指、质量最差、低质量、jpeg伪影、模糊、糟糕,黑脸,霓虹灯" ) 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=30, ) with gr.Row(): controlnet_conditioning_scale = gr.Slider( label="Controlnet Conditioning Scale", minimum=0.0, maximum=1.0, step=0.1, value=0.5, ) control_guidance_end = gr.Slider( label="Control Guidance End", minimum=0.0, maximum=1.0, step=0.1, value=0.9, ) with gr.Row(): strength = gr.Slider( label="Strength", minimum=0.0, maximum=1.0, step=0.1, value=1.0, ) ip_scale = gr.Slider( label="IP_Scale", minimum=0.0, maximum=1.0, step=0.1, value=0.5, ) with gr.Row(): canny_button = gr.Button("Canny", elem_id="button") depth_button = gr.Button("Depth", elem_id="button") pose_button = gr.Button("Pose", elem_id="button") with gr.Column(elem_id="col-right"): result = gr.Gallery(label="Result", show_label=False, columns=2) seed_used = gr.Number(label="Seed Used") with gr.Row(): gr.Examples( fn = infer_canny, examples = canny_examples, inputs = [prompt, image, ipa_image], outputs = [result, seed_used], label = "Canny" ) with gr.Row(): gr.Examples( fn = infer_depth, examples = depth_examples, inputs = [prompt, image, ipa_image], outputs = [result, seed_used], label = "Depth" ) with gr.Row(): gr.Examples( fn = infer_pose, examples = pose_examples, inputs = [prompt, image, ipa_image], outputs = [result, seed_used], label = "Pose" ) canny_button.click( fn = infer_canny, inputs = [prompt, image, ipa_image, negative_prompt, seed, randomize_seed, guidance_scale, num_inference_steps, controlnet_conditioning_scale, control_guidance_end, strength, ip_scale], outputs = [result, seed_used] ) depth_button.click( fn = infer_depth, inputs = [prompt, image, ipa_image, negative_prompt, seed, randomize_seed, guidance_scale, num_inference_steps, controlnet_conditioning_scale, control_guidance_end, strength, ip_scale], outputs = [result, seed_used] ) pose_button.click( fn = infer_pose, inputs = [prompt, image, ipa_image, negative_prompt, seed, randomize_seed, guidance_scale, num_inference_steps, controlnet_conditioning_scale, control_guidance_end, strength, ip_scale], outputs = [result, seed_used] ) Kolors.queue().launch(debug=True)