import os import gradio as gr from gradio_imageslider import ImageSlider import argparse from SUPIR.util import HWC3, upscale_image, fix_resize, convert_dtype import numpy as np import torch from SUPIR.util import create_SUPIR_model, load_QF_ckpt from PIL import Image from llava.llava_agent import LLavaAgent from CKPT_PTH import LLAVA_MODEL_PATH import einops import copy import math import time import random import spaces from huggingface_hub import hf_hub_download hf_hub_download(repo_id="laion/CLIP-ViT-bigG-14-laion2B-39B-b160k", filename="open_clip_pytorch_model.bin", local_dir="laion_CLIP-ViT-bigG-14-laion2B-39B-b160k") hf_hub_download(repo_id="camenduru/SUPIR", filename="sd_xl_base_1.0_0.9vae.safetensors", local_dir="yushan777_SUPIR") hf_hub_download(repo_id="camenduru/SUPIR", filename="SUPIR-v0F.ckpt", local_dir="yushan777_SUPIR") hf_hub_download(repo_id="camenduru/SUPIR", filename="SUPIR-v0Q.ckpt", local_dir="yushan777_SUPIR") hf_hub_download(repo_id="RunDiffusion/Juggernaut-XL-Lightning", filename="Juggernaut_RunDiffusionPhoto2_Lightning_4Steps.safetensors", local_dir="RunDiffusion_Juggernaut-XL-Lightning") parser = argparse.ArgumentParser() parser.add_argument("--opt", type=str, default='options/SUPIR_v0.yaml') parser.add_argument("--ip", type=str, default='127.0.0.1') parser.add_argument("--port", type=int, default='6688') parser.add_argument("--no_llava", action='store_true', default=True)#False parser.add_argument("--use_image_slider", action='store_true', default=False)#False parser.add_argument("--log_history", action='store_true', default=False) parser.add_argument("--loading_half_params", action='store_true', default=False)#False parser.add_argument("--use_tile_vae", action='store_true', default=True)#False parser.add_argument("--encoder_tile_size", type=int, default=512) parser.add_argument("--decoder_tile_size", type=int, default=64) parser.add_argument("--load_8bit_llava", action='store_true', default=False) args = parser.parse_args() use_llava = not args.no_llava if torch.cuda.device_count() > 0: if torch.cuda.device_count() >= 2: SUPIR_device = 'cuda:0' LLaVA_device = 'cuda:1' elif torch.cuda.device_count() == 1: SUPIR_device = 'cuda:0' LLaVA_device = 'cuda:0' else: SUPIR_device = 'cpu' LLaVA_device = 'cpu' # load SUPIR model, default_setting = create_SUPIR_model(args.opt, SUPIR_sign='Q', load_default_setting=True) if args.loading_half_params: model = model.half() if args.use_tile_vae: model.init_tile_vae(encoder_tile_size=args.encoder_tile_size, decoder_tile_size=args.decoder_tile_size) model = model.to(SUPIR_device) model.first_stage_model.denoise_encoder_s1 = copy.deepcopy(model.first_stage_model.denoise_encoder) model.current_model = 'v0-Q' ckpt_Q, ckpt_F = load_QF_ckpt(args.opt) # load LLaVA if use_llava: llava_agent = LLavaAgent(LLAVA_MODEL_PATH, device=LLaVA_device, load_8bit=args.load_8bit_llava, load_4bit=False) else: llava_agent = None def update_seed(is_randomize_seed, seed): if is_randomize_seed: return random.randint(0, 2147483647) return seed def check(input_image): if input_image is None: raise gr.Error("Please provide an image to restore.") def reset_feedback(): return 3, '' @spaces.GPU(duration=540) def stage1_process(input_image, gamma_correction): print('stage1_process ==>>') if torch.cuda.device_count() == 0: gr.Warning('Set this space to GPU config to make it work.') return None, None torch.cuda.set_device(SUPIR_device) LQ = HWC3(input_image) LQ = fix_resize(LQ, 512) # stage1 LQ = np.array(LQ) / 255 * 2 - 1 LQ = torch.tensor(LQ, dtype=torch.float32).permute(2, 0, 1).unsqueeze(0).to(SUPIR_device)[:, :3, :, :] LQ = model.batchify_denoise(LQ, is_stage1=True) LQ = (LQ[0].permute(1, 2, 0) * 127.5 + 127.5).cpu().numpy().round().clip(0, 255).astype(np.uint8) # gamma correction LQ = LQ / 255.0 LQ = np.power(LQ, gamma_correction) LQ *= 255.0 LQ = LQ.round().clip(0, 255).astype(np.uint8) print('<<== stage1_process') return LQ, gr.update(visible = True) @spaces.GPU(duration=540) def llave_process(input_image, temperature, top_p, qs=None): print('llave_process ==>>') if torch.cuda.device_count() == 0: gr.Warning('Set this space to GPU config to make it work.') return 'Set this space to GPU config to make it work.' torch.cuda.set_device(LLaVA_device) if use_llava: LQ = HWC3(input_image) LQ = Image.fromarray(LQ.astype('uint8')) captions = llava_agent.gen_image_caption([LQ], temperature=temperature, top_p=top_p, qs=qs) else: captions = ['LLaVA is not available. Please add text manually.'] print('<<== llave_process') return captions[0] @spaces.GPU(duration=540) def stage2_process( noisy_image, denoise_image, prompt, a_prompt, n_prompt, num_samples, min_size, downscale, upscale, edm_steps, s_stage1, s_stage2, s_cfg, randomize_seed, seed, s_churn, s_noise, color_fix_type, diff_dtype, ae_dtype, gamma_correction, linear_CFG, linear_s_stage2, spt_linear_CFG, spt_linear_s_stage2, model_select, output_format ): start = time.time() print('stage2_process ==>>') if torch.cuda.device_count() == 0: gr.Warning('Set this space to GPU config to make it work.') return None, None, None if output_format == "input": if noisy_image is None: output_format = "png" else: output_format = noisy_image.format input_image = noisy_image if denoise_image is None else denoise_image if 1 < downscale: input_height, input_width, input_channel = np.array(input_image).shape input_image = input_image.resize((input_width // downscale, input_height // downscale), Image.LANCZOS) torch.cuda.set_device(SUPIR_device) event_id = str(time.time_ns()) event_dict = {'event_id': event_id, 'localtime': time.ctime(), 'prompt': prompt, 'a_prompt': a_prompt, 'n_prompt': n_prompt, 'num_samples': num_samples, 'upscale': upscale, 'edm_steps': edm_steps, 's_stage1': s_stage1, 's_stage2': s_stage2, 's_cfg': s_cfg, 'seed': seed, 's_churn': s_churn, 's_noise': s_noise, 'color_fix_type': color_fix_type, 'diff_dtype': diff_dtype, 'ae_dtype': ae_dtype, 'gamma_correction': gamma_correction, 'linear_CFG': linear_CFG, 'linear_s_stage2': linear_s_stage2, 'spt_linear_CFG': spt_linear_CFG, 'spt_linear_s_stage2': spt_linear_s_stage2, 'model_select': model_select} if model_select != model.current_model: print('load ' + model_select) if model_select == 'v0-Q': model.load_state_dict(ckpt_Q, strict=False) elif model_select == 'v0-F': model.load_state_dict(ckpt_F, strict=False) model.current_model = model_select input_image = HWC3(input_image) input_image = upscale_image(input_image, upscale, unit_resolution=32, min_size=min_size) LQ = np.array(input_image) / 255.0 LQ = np.power(LQ, gamma_correction) LQ *= 255.0 LQ = LQ.round().clip(0, 255).astype(np.uint8) LQ = LQ / 255 * 2 - 1 LQ = torch.tensor(LQ, dtype=torch.float32).permute(2, 0, 1).unsqueeze(0).to(SUPIR_device)[:, :3, :, :] if use_llava: captions = [prompt] else: captions = [''] model.ae_dtype = convert_dtype(ae_dtype) model.model.dtype = convert_dtype(diff_dtype) samples = model.batchify_sample(LQ, captions, num_steps=edm_steps, restoration_scale=s_stage1, s_churn=s_churn, s_noise=s_noise, cfg_scale=s_cfg, control_scale=s_stage2, seed=seed, num_samples=num_samples, p_p=a_prompt, n_p=n_prompt, color_fix_type=color_fix_type, use_linear_CFG=linear_CFG, use_linear_control_scale=linear_s_stage2, cfg_scale_start=spt_linear_CFG, control_scale_start=spt_linear_s_stage2) x_samples = (einops.rearrange(samples, 'b c h w -> b h w c') * 127.5 + 127.5).cpu().numpy().round().clip( 0, 255).astype(np.uint8) results = [x_samples[i] for i in range(num_samples)] if args.log_history: os.makedirs(f'./history/{event_id[:5]}/{event_id[5:]}', exist_ok=True) with open(f'./history/{event_id[:5]}/{event_id[5:]}/logs.txt', 'w') as f: f.write(str(event_dict)) f.close() Image.fromarray(input_image).save(f'./history/{event_id[:5]}/{event_id[5:]}/LQ.png') for i, result in enumerate(results): Image.fromarray(result).save(f'./history/{event_id[:5]}/{event_id[5:]}/HQ_{i}.png') # All the results have the same size result_height, result_width, result_channel = np.array(results[0]).shape print('<<== stage2_process') end = time.time() secondes = int(end - start) minutes = math.floor(secondes / 60) secondes = secondes - (minutes * 60) hours = math.floor(minutes / 60) minutes = minutes - (hours * 60) information = ("Start the process again if you want a different result. " if randomize_seed else "") + \ "The new image resolution is " + str(result_width) + \ " pixels large and " + str(result_height) + \ " pixels high, so a resolution of " + f'{result_width * result_height:,}' + " pixels. " + \ "The image(s) has(ve) been generated in " + \ ((str(hours) + " h, ") if hours != 0 else "") + \ ((str(minutes) + " min, ") if hours != 0 or minutes != 0 else "") + \ str(secondes) + " sec." print(information) # Only one image can be shown in the slider return [noisy_image] + [results[0]], gr.update(format = output_format, value = [noisy_image] + results), gr.update(value = information, visible = True), event_id def load_and_reset(param_setting): print('load_and_reset ==>>') if torch.cuda.device_count() == 0: gr.Warning('Set this space to GPU config to make it work.') return None, None, None, None, None, None, None, None, None, None, None, None, None, None edm_steps = default_setting.edm_steps s_stage2 = 1.0 s_stage1 = -1.0 s_churn = 5 s_noise = 1.003 a_prompt = 'Cinematic, High Contrast, highly detailed, taken using a Canon EOS R camera, hyper detailed photo - ' \ 'realistic maximum detail, 32k, Color Grading, ultra HD, extreme meticulous detailing, skin pore ' \ 'detailing, hyper sharpness, perfect without deformations.' n_prompt = 'painting, oil painting, illustration, drawing, art, sketch, anime, cartoon, CG Style, ' \ '3D render, unreal engine, blurring, dirty, messy, worst quality, low quality, frames, watermark, ' \ 'signature, jpeg artifacts, deformed, lowres, over-smooth' color_fix_type = 'Wavelet' spt_linear_s_stage2 = 0.0 linear_s_stage2 = False linear_CFG = True if param_setting == "Quality": s_cfg = default_setting.s_cfg_Quality spt_linear_CFG = default_setting.spt_linear_CFG_Quality model_select = "v0-Q" elif param_setting == "Fidelity": s_cfg = default_setting.s_cfg_Fidelity spt_linear_CFG = default_setting.spt_linear_CFG_Fidelity model_select = "v0-F" else: raise NotImplementedError gr.Info('The parameters are reset.') print('<<== load_and_reset') return edm_steps, s_cfg, s_stage2, s_stage1, s_churn, s_noise, a_prompt, n_prompt, color_fix_type, linear_CFG, \ linear_s_stage2, spt_linear_CFG, spt_linear_s_stage2, model_select def submit_feedback(event_id, fb_score, fb_text): if args.log_history: with open(f'./history/{event_id[:5]}/{event_id[5:]}/logs.txt', 'r') as f: event_dict = eval(f.read()) f.close() event_dict['feedback'] = {'score': fb_score, 'text': fb_text} with open(f'./history/{event_id[:5]}/{event_id[5:]}/logs.txt', 'w') as f: f.write(str(event_dict)) f.close() return 'Submit successfully, thank you for your comments!' else: return 'Submit failed, the server is not set to log history.' title_html = """

SUPIR

Upscale your images up to x8 freely, without account, without watermark and download it
🤸🤸

This is an online demo of SUPIR, a practicing model scaling for photo-realistic image restoration. It is still a research project under tested and is not yet a stable commercial product. LlaVa is not integrated in this demo. The content added by SUPIR is imagination, not real-world information. The aim of SUPIR is the beauty and the illustration. Most of the processes only last few minutes. This demo can handle huge images but the process will be aborted if it lasts more than 9 min.

Paper   Project Page   How to play   Local Install Guide

""" claim_md = """ ## **Piracy** The images are not stored but the logs are saved during a month. ## **Terms of use** By using this service, users are required to agree to the following terms: The service is a research preview intended for non-commercial use only. It only provides limited safety measures and may generate offensive content. It must not be used for any illegal, harmful, violent, racist, or sexual purposes. The service may collect user dialogue data for future research. Please submit a feedback to us if you get any inappropriate answer! We will collect those to keep improving our models. For an optimal experience, please use desktop computers for this demo, as mobile devices may compromise its quality. ## **License** The service is a research preview intended for non-commercial use only, subject to the model [License](https://github.com/Fanghua-Yu/SUPIR) of SUPIR. """ # Gradio interface with gr.Blocks(title="SUPIR") as interface: if torch.cuda.device_count() == 0: with gr.Row(): gr.HTML("""

⚠️To use SUPIR, Duplicate this space and set a GPU with 30 GB VRAM. You can't use SUPIR directly here because this space runs on a CPU, which is not enough for SUPIR. This is a template space. Please provide feedback if you have issues.

""") gr.HTML(title_html) input_image = gr.Image(label="Input", show_label=True, type="numpy", height=600, elem_id="image-input") with gr.Group(): prompt = gr.Textbox(label="Image description for LlaVa", value="", placeholder="A person, walking, in a town, Summer, photorealistic", lines=3, visible=False) a_prompt = gr.Textbox(label="Image description", info="Help the AI understand what the image represents; describe as much as possible", value='Cinematic, High Contrast, highly detailed, taken using a Canon EOS R ' 'camera, hyper detailed photo - realistic maximum detail, 32k, Color ' 'Grading, ultra HD, extreme meticulous detailing, skin pore detailing, ' 'hyper sharpness, perfect without deformations.', lines=3) a_prompt_hint = gr.HTML("You can use a LlaVa space to auto-generate the description of your image.") upscale = gr.Radio([["x1", 1], ["x2", 2], ["x3", 3], ["x4", 4], ["x5", 5], ["x6", 6], ["x7", 7], ["x8", 8]], label="Upscale factor", info="Resolution x1 to x8", value=2, interactive=True) output_format = gr.Radio([["*.png", "png"], ["*.webp", "webp"], ["*.jpeg", "jpeg"], ["*.gif", "gif"], ["*.bmp", "bmp"]], label="Image format for result", info="File extention", value="png", interactive=True) with gr.Accordion("Pre-denoising (optional)", open=False): gamma_correction = gr.Slider(label="Gamma Correction", info = "lower=lighter, higher=darker", minimum=0.1, maximum=2.0, value=1.0, step=0.1) denoise_button = gr.Button(value="Pre-denoise") denoise_image = gr.Image(label="Denoised image", show_label=True, type="numpy", height=600, elem_id="image-s1") denoise_information = gr.HTML(value="If present, the denoised image will be used for the restoration instead of the input image.", visible=False) with gr.Accordion("LLaVA options", open=False, visible=False): temperature = gr.Slider(label="Temperature", info = "lower=Always similar, higher=More creative", minimum=0., maximum=1.0, value=0.2, step=0.1) top_p = gr.Slider(label="Top P", info = "Percent of tokens shortlisted", minimum=0., maximum=1.0, value=0.7, step=0.1) qs = gr.Textbox(label="Question", info="Ask LLaVa what description you want", value="Describe the image and its style in a very detailed manner. The image is a realistic photography, not an art painting.", lines=3) with gr.Accordion("Advanced options", open=False): n_prompt = gr.Textbox(label="Anti image description", info="Disambiguate by listing what the image does NOT represent", value='painting, oil painting, illustration, drawing, art, sketch, anime, ' 'cartoon, CG Style, 3D render, unreal engine, blurring, bokeh, ugly, dirty, messy, ' 'worst quality, low quality, frames, watermark, signature, jpeg artifacts, ' 'deformed, lowres, over-smooth', lines=3) edm_steps = gr.Slider(label="Steps", info="lower=faster, higher=more details", minimum=1, maximum=200, value=default_setting.edm_steps if torch.cuda.device_count() > 0 else 1, step=1) num_samples = gr.Slider(label="Num Samples", info="Number of generated results", minimum=1, maximum=4 if not args.use_image_slider else 1 , value=1, step=1) min_size = gr.Slider(label="Minimum size", info="Minimum height, minimum width of the result", minimum=32, maximum=4096, value=1024, step=32) downscale = gr.Radio([["/1", 1], ["/2", 2], ["/3", 3], ["/4", 4], ["/5", 5], ["/6", 6], ["/7", 7], ["/8", 8]], label="Pre-downscale factor", info="Reducing blurred image reduce the process time", value=1, interactive=True) with gr.Row(): with gr.Column(): model_select = gr.Radio([["💃 Quality", "v0-Q"], ["🎯 Fidelity", "v0-F"]], label="Model Selection", info="Pretrained model", value="v0-Q", interactive=True) with gr.Column(): color_fix_type = gr.Radio(["None", "AdaIn", "Wavelet"], label="Color-Fix Type", info="AdaIn=Improve following a style, Wavelet=For JPEG artifacts", value="Wavelet", interactive=True) s_cfg = gr.Slider(label="Text Guidance Scale", info="lower=follow the image, higher=follow the prompt", minimum=1.0, maximum=15.0, value=default_setting.s_cfg_Quality if torch.cuda.device_count() > 0 else 1.0, step=0.1) s_stage2 = gr.Slider(label="Restoring Guidance Strength", minimum=0., maximum=1., value=1., step=0.05) s_stage1 = gr.Slider(label="Pre-denoising Guidance Strength", minimum=-1.0, maximum=6.0, value=-1.0, step=1.0) s_churn = gr.Slider(label="S-Churn", minimum=0, maximum=40, value=5, step=1) s_noise = gr.Slider(label="S-Noise", minimum=1.0, maximum=1.1, value=1.003, step=0.001) with gr.Row(): with gr.Column(): linear_CFG = gr.Checkbox(label="Linear CFG", value=True) spt_linear_CFG = gr.Slider(label="CFG Start", minimum=1.0, maximum=9.0, value=default_setting.spt_linear_CFG_Quality if torch.cuda.device_count() > 0 else 1.0, step=0.5) with gr.Column(): linear_s_stage2 = gr.Checkbox(label="Linear Restoring Guidance", value=False) spt_linear_s_stage2 = gr.Slider(label="Guidance Start", minimum=0., maximum=1., value=0., step=0.05) with gr.Column(): diff_dtype = gr.Radio(['fp32', 'fp16', 'bf16'], label="Diffusion Data Type", value="fp16", interactive=True) with gr.Column(): ae_dtype = gr.Radio(['fp32', 'bf16'], label="Auto-Encoder Data Type", value="bf16", interactive=True) randomize_seed = gr.Checkbox(label = "\U0001F3B2 Randomize seed", value = True, info = "If checked, result is always different") seed = gr.Slider(label="Seed", minimum=0, maximum=2147483647, step=1, randomize=True) with gr.Group(): param_setting = gr.Radio(["Quality", "Fidelity"], interactive=True, label="Presetting", value="Quality") restart_button = gr.Button(value="Apply presetting") with gr.Group(): llave_button = gr.Button(value="Generate description by LlaVa (disabled)", visible=False) diffusion_button = gr.Button(value="🚀 Upscale/Restore", variant = "primary", elem_id="process_button") restore_information = gr.HTML(value="Restart the process to get another result.", visible=False) result_slider = ImageSlider(label='Output', show_label=True, elem_id="slider1") result_gallery = gr.Gallery(label='Output', show_label=True, elem_id="gallery1") with gr.Accordion("Feedback", open=True, visible=False): fb_score = gr.Slider(label="Feedback Score", minimum=1, maximum=5, value=3, step=1, interactive=True) fb_text = gr.Textbox(label="Feedback Text", value="", placeholder='Please enter your feedback here.') submit_button = gr.Button(value="Submit Feedback") event_id = gr.Textbox(label="Event ID", value="", visible=False) gr.Examples( fn = stage2_process, inputs = [ input_image, denoise_image, prompt, a_prompt, n_prompt, num_samples, min_size, downscale, upscale, edm_steps, s_stage1, s_stage2, s_cfg, randomize_seed, seed, s_churn, s_noise, color_fix_type, diff_dtype, ae_dtype, gamma_correction, linear_CFG, linear_s_stage2, spt_linear_CFG, spt_linear_s_stage2, model_select, output_format ], outputs = [ result_slider, result_gallery, restore_information, event_id ], examples = [ [ "./Examples/Example1.png", None, None, "Group of people, walking, happy, in the street, photorealistic, 8k, extremely detailled, Cinematic, High Contrast, highly detailed, taken using a Canon EOS R camera, hyper detailed photo - realistic maximum detail, 32k, Color Grading, ultra HD, extreme meticulous detailing, skin pore detailing, hyper sharpness, perfect without deformations.", "painting, oil painting, illustration, drawing, art, sketch, anime, cartoon, CG Style, 3D render, unreal engine, blurring, bokeh, ugly, dirty, messy, worst quality, low quality, frames, watermark, signature, jpeg artifacts, deformed, lowres, over-smooth", 1, 1024, 1, 1, 200, -1, 1, 7.5, False, 42, 5, 1.003, "Wavelet", "fp16", "bf16", 1.0, True, 4, False, 0., "v0-Q", "png" ], [ "./Examples/Example2.jpeg", None, None, "The head of a tabby cat, in a house, photorealistic, 8k, extremely detailled, Cinematic, High Contrast, highly detailed, taken using a Canon EOS R camera, hyper detailed photo - realistic maximum detail, 32k, Color Grading, ultra HD, extreme meticulous detailing, skin pore detailing, hyper sharpness, perfect without deformations.", "painting, oil painting, illustration, drawing, art, sketch, anime, cartoon, CG Style, 3D render, unreal engine, blurring, bokeh, ugly, dirty, messy, worst quality, low quality, frames, watermark, signature, jpeg artifacts, deformed, lowres, over-smooth", 1, 1024, 1, 1, 200, -1, 1, 7.5, False, 42, 5, 1.003, "Wavelet", "fp16", "bf16", 1.0, True, 4, False, 0., "v0-Q", "png" ], ], cache_examples = False, ) with gr.Row(): gr.Markdown(claim_md) denoise_button.click(fn = check, inputs = [ input_image ], outputs = [], queue = False, show_progress = False).success(fn = stage1_process, inputs = [ input_image, gamma_correction ], outputs=[ denoise_image, denoise_information ]) llave_button.click(fn = check, inputs = [ denoise_image ], outputs = [], queue = False, show_progress = False).success(fn = llave_process, inputs = [ denoise_image, temperature, top_p, qs ], outputs = [ prompt ]) diffusion_button.click(fn = update_seed, inputs = [ randomize_seed, seed ], outputs = [ seed ], queue = False, show_progress = False).then(fn = check, inputs = [ input_image ], outputs = [], queue = False, show_progress = False).success(fn = reset_feedback, inputs = [], outputs = [ fb_score, fb_text ], queue = False, show_progress = False).success(fn=stage2_process, inputs = [ input_image, denoise_image, prompt, a_prompt, n_prompt, num_samples, min_size, downscale, upscale, edm_steps, s_stage1, s_stage2, s_cfg, randomize_seed, seed, s_churn, s_noise, color_fix_type, diff_dtype, ae_dtype, gamma_correction, linear_CFG, linear_s_stage2, spt_linear_CFG, spt_linear_s_stage2, model_select, output_format ], outputs = [ result_slider, result_gallery, restore_information, event_id ]) restart_button.click(fn = load_and_reset, inputs = [ param_setting ], outputs = [ edm_steps, s_cfg, s_stage2, s_stage1, s_churn, s_noise, a_prompt, n_prompt, color_fix_type, linear_CFG, linear_s_stage2, spt_linear_CFG, spt_linear_s_stage2, model_select ]) submit_button.click(fn = submit_feedback, inputs = [ event_id, fb_score, fb_text ], outputs = [ fb_text ]) interface.queue(10).launch()