import os import gradio as gr import argparse import numpy as np import torch import einops import copy import math import time import random import spaces import re import uuid from PIL import Image from SUPIR.util import HWC3, upscale_image, fix_resize, convert_dtype, create_SUPIR_model, load_QF_ckpt from huggingface_hub import hf_hub_download from pillow_heif import register_heif_opener register_heif_opener() max_64_bit_int = np.iinfo(np.int32).max from llava.llava_agent import LLavaAgent from CKPT_PTH import LLAVA_MODEL_PATH 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=True) # False parser.add_argument("--use_tile_vae", action='store_true', default=False) # 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=True) args = parser.parse_args() use_llava = not args.no_llava if torch.cuda.device_count() >= 2: SUPIR_device = 'cuda:0' LLaVA_device = 'cuda:1' # 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) elif torch.cuda.device_count() == 1: SUPIR_device = 'cuda:0' LLaVA_device = 'cuda:0' # 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) else: raise ValueError('Currently support CUDA only.') # 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 check_upload(input_image): if input_image is None: raise gr.Error("Please provide an image to restore.") return gr.update(visible=True) def update_seed(is_randomize_seed, seed): if is_randomize_seed: return random.randint(0, max_64_bit_int) return seed def reset(): return [ None, 0, None, None, "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, aliasing, unsharp, weird textures, ugly, dirty, messy, worst quality, low quality, frames, watermark, signature, jpeg artifacts, deformed, lowres, over-smooth", 1, 1024, 1, 2, 50, -1.0, 1.0, default_setting.s_cfg_Quality if torch.cuda.device_count() > 0 else 1.0, True, random.randint(0, max_64_bit_int), 5, 1.003, "Wavelet", "fp32", "fp32", 1.0, True, False, default_setting.spt_linear_CFG_Quality if torch.cuda.device_count() > 0 else 1.0, 0.0, "v0-Q", "input", 6 ] def check(input_image): if input_image is None: raise gr.Error("Please provide an image to restore.") @spaces.GPU(duration=20) def stage1_process( input_image, gamma_correction, diff_dtype, ae_dtype ): 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(np.array(Image.open(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, :, :] model.ae_dtype = convert_dtype(ae_dtype) model.model.dtype = convert_dtype(diff_dtype) 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 = np.power(LQ, float(gamma_correction) if isinstance(gamma_correction, (int, float)) else gamma_correction.value) LQ *= 255.0 LQ = LQ.round().clip(0, 255).astype(np.uint8) print('<<== stage1_process') return LQ, gr.update(visible=True) def stage2_process(*args, **kwargs): try: return restore_in_Xmin(*args, **kwargs) except Exception as e: print("An error occurred during processing:", str(e)) raise e def restore_in_Xmin( noisy_image, rotation, 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, allocation ): print("Starting restoration process...") input_format = re.sub(r"^.*\.([^\.]+)$", r"\1", noisy_image) if input_format not in ['png', 'webp', 'jpg', 'jpeg', 'gif', 'bmp', 'heic']: gr.Warning('Invalid image format. Please first convert into *.png, *.webp, *.jpg, *.jpeg, *.gif, *.bmp or *.heic.') return None, gr.update(value="Invalid image format.", visible=True), gr.update(visible=True) if output_format == "input": if noisy_image is None: output_format = "png" else: output_format = input_format print("Final output_format:", output_format) if prompt is None: prompt = "" if a_prompt is None: a_prompt = "" if n_prompt is None: n_prompt = "" if prompt != "" and a_prompt != "": a_prompt = prompt + ", " + a_prompt else: a_prompt = prompt + a_prompt print("Final prompt:", a_prompt) denoise_image = np.array(Image.open(noisy_image if denoise_image is None else denoise_image)) if rotation == 90: denoise_image = np.array(list(zip(*denoise_image[::-1]))) elif rotation == 180: denoise_image = np.array(list(zip(*denoise_image[::-1]))) denoise_image = np.array(list(zip(*denoise_image[::-1]))) elif rotation == -90: denoise_image = np.array(list(zip(*denoise_image))[::-1]) if 1 < downscale: input_height, input_width, input_channel = denoise_image.shape denoise_image = np.array(Image.fromarray(denoise_image).resize((input_width // downscale, input_height // downscale), Image.LANCZOS)) denoise_image = HWC3(denoise_image) if torch.cuda.device_count() == 0: gr.Warning('Set this space to GPU config to make it work.') return denoise_image, gr.update(value="No GPU available.", visible=True), gr.update(visible=True) if model_select != model.current_model: print('Loading model:', 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 model.ae_dtype = convert_dtype(ae_dtype) model.model.dtype = convert_dtype(diff_dtype) # Allocation if allocation == 1: return restore_in_1min( noisy_image, denoise_image, a_prompt, n_prompt, num_samples, min_size, upscale, edm_steps, s_stage1, s_stage2, s_cfg, seed, s_churn, s_noise, color_fix_type, linear_CFG, linear_s_stage2, spt_linear_CFG, spt_linear_s_stage2, output_format ) # 他のallocation条件は省略 # デフォルトの処理 return restore_in_1min( noisy_image, denoise_image, a_prompt, n_prompt, num_samples, min_size, upscale, edm_steps, s_stage1, s_stage2, s_cfg, seed, s_churn, s_noise, color_fix_type, linear_CFG, linear_s_stage2, spt_linear_CFG, spt_linear_s_stage2, output_format ) @spaces.GPU(duration=89) def restore_in_1min(*args, **kwargs): return restore_on_gpu(*args, **kwargs) def restore_on_gpu( noisy_image, input_image, a_prompt, n_prompt, num_samples, min_size, upscale, edm_steps, s_stage1, s_stage2, s_cfg, seed, s_churn, s_noise, color_fix_type, linear_CFG, linear_s_stage2, spt_linear_CFG, spt_linear_s_stage2, output_format ): start = time.time() print('Starting GPU restoration...') torch.cuda.set_device(SUPIR_device) with torch.no_grad(): 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 = np.power(LQ, float(gamma_correction) if isinstance(gamma_correction, (int, float)) else gamma_correction.value) 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, :, :] captions = [''] 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)] torch.cuda.empty_cache() # 結果の処理 result_image = results[0] result_pil = Image.fromarray(result_image) end = time.time() elapsed_time = end - start information = f"Processing completed in {elapsed_time:.2f} seconds." print(information) print('GPU restoration completed.') return result_pil, gr.update(value=information, visible=True), gr.update(visible=True) 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 title_html = """

SUPIR

Upscale your images up to x10 freely, without account, without watermark and download it
🤸🤸

This is an online demo of SUPIR, a practicing model scaling for photo-realistic image restoration. The content added by SUPIR is imagination, not real-world information. SUPIR is for beauty and illustration only. Most of the processes last few minutes. If you want to upscale AI-generated images, be noticed that PixArt Sigma space can directly generate 5984x5984 images. Due to Gradio issues, the generated image is slightly less saturated than the original. Please leave a message in discussion if you encounter issues. You can also use AuraSR to upscale x4.

Paper   Project Page   Local Install Guide

GitHub Repo stars

""" claim_md = """ ## **Piracy** The images are not stored but the logs are saved during a month. ## **How to get SUPIR** You can get SUPIR on HuggingFace by [duplicating this space](https://huggingface.co/spaces/Fabrice-TIERCELIN/SUPIR?duplicate=true) and set GPU. You can also install SUPIR on your computer following [this tutorial](https://huggingface.co/blog/MonsterMMORPG/supir-sota-image-upscale-better-than-magnific-ai). You can install _Pinokio_ on your computer and then install _SUPIR_ into it. It should be quite easy if you have an Nvidia GPU. ## **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() 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. Please provide feedback if you have issues.

""") gr.HTML(title_html) input_image = gr.Image(label="Input (*.png, *.webp, *.jpeg, *.jpg, *.gif, *.bmp, *.heic)", show_label=True, type="filepath", height=600, elem_id="image-input") rotation = gr.Radio([["No rotation", 0], ["⤵ Rotate +90°", 90], ["↩ Return 180°", 180], ["⤴ Rotate -90°", -90]], label="Orientation correction", info="Will apply the following rotation before restoring the image; the AI needs a good orientation to understand the content", value=0, interactive=True, visible=False) with gr.Group(): prompt = gr.Textbox(label="Image description", info="Help the AI understand what the image represents; describe as much as possible, especially the details we can't see on the original image; you can write in any language", value="", placeholder="A 33 years old man, walking, in the street, Santiago, morning, Summer, photorealistic", lines=3) 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], ["x9", 9], ["x10", 10]], label="Upscale factor", info="Resolution x1 to x10", value=2, interactive=True) output_format = gr.Radio([["As input", "input"], ["*.png", "png"], ["*.webp", "webp"], ["*.jpeg", "jpeg"], ["*.gif", "gif"], ["*.bmp", "bmp"]], label="Image format for result", info="File extension", value="input", interactive=True) allocation = gr.Radio([["1 min", 1], ["2 min", 2], ["3 min", 3], ["4 min", 4], ["5 min", 5]], label="GPU allocation time", info="lower=May abort run, higher=Quota penalty for next runs", value=1, 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="filepath", sources=[], interactive=False, 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("Advanced options", open=False): a_prompt = gr.Textbox(label="Additional image description", info="Completes the main image description", 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) n_prompt = gr.Textbox(label="Negative 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, aliasing, unsharp, weird textures, 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=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], ["/9", 9], ["/10", 10]], 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)", "v0-Q"], ["🎯 Fidelity (v0-F)", "v0-F"]], label="Model Selection", info="Pretrained model", value="v0-Q", interactive=True) with gr.Column(): color_fix_type = gr.Radio([["None", "None"], ["AdaIn (improve as a photo)", "AdaIn"], ["Wavelet (for JPEG artifacts)", "Wavelet"]], label="Color-Fix Type", info="AdaIn=Improve following a style, Wavelet=For JPEG artifacts", value="AdaIn", 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=0.0, maximum=1.0, value=1.0, 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=0.0, maximum=1.0, value=0.0, step=0.05) with gr.Column(): diff_dtype = gr.Radio([["fp32 (precision)", "fp32"], ["fp16 (medium)", "fp16"], ["bf16 (speed)", "bf16"]], label="Diffusion Data Type", value="fp32", interactive=True) with gr.Column(): ae_dtype = gr.Radio([["fp32 (precision)", "fp32"], ["bf16 (speed)", "bf16"]], label="Auto-Encoder Data Type", value="fp32", 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=max_64_bit_int, 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.Column(): diffusion_button = gr.Button(value="🚀 Upscale/Restore", variant="primary", elem_id="process_button") reset_btn = gr.Button(value="🧹 Reinit page", variant="stop", elem_id="reset_button", visible=False) restore_information = gr.HTML(value="Restart the process to get another result.", visible=False) result_image = gr.Image(label='Result Image', show_label=True, interactive=False, elem_id='result_image') gr.Examples( examples=[ # 例を適宜追加または削除 ], run_on_click=True, fn=stage2_process, inputs=[ input_image, rotation, 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, allocation ], outputs=[ result_image, restore_information, reset_btn ], cache_examples=False, ) with gr.Row(): gr.Markdown(claim_md) input_image.upload(fn=check_upload, inputs=[ input_image ], outputs=[ rotation ], queue=False, show_progress=False) denoise_button.click(fn=check, inputs=[ input_image ], outputs=[], queue=False, show_progress=False).success(fn=stage1_process, inputs=[ input_image, gamma_correction, diff_dtype, ae_dtype ], outputs=[ denoise_image, denoise_information ]) 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=stage2_process, inputs=[ input_image, rotation, 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, allocation ], outputs=[ result_image, restore_information, reset_btn ]) 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 ]) reset_btn.click(fn=reset, inputs=[], outputs=[ input_image, rotation, 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, allocation ], queue=False, show_progress=False) interface.queue(10).launch(show_error=True)