Scaling Up to Excellence: Practicing Model Scaling for Photo-Realistic Image Restoration In the Wild
[Paper] [Project Page] [Replicate Demo]
Fanghua, Yu, Jinjin Gu, Zheyuan Li, Jinfan Hu, Xiangtao Kong, Xintao Wang, Jingwen He, Yu Qiao, Chao Dong
Shenzhen Institute of Advanced Technology; Shanghai AI Laboratory; University of Sydney; The Hong Kong Polytechnic University; ARC Lab, Tencent PCG; The Chinese University of Hong Kong
-
Clone repo
git clone https://github.com/Fanghua-Yu/SUPIR.git cd SUPIR -
Install dependent packages
conda create -n SUPIR python=3.8 -y conda activate SUPIR pip install --upgrade pip pip install -r requirements.txt
-
Download Checkpoints
For users who can connect to huggingface, please setting LLAVA_CLIP_PATH, SDXL_CLIP1_PATH, SDXL_CLIP2_CKPT_PTH in CKPT_PTH.py as None. These CLIPs will be downloaded automatically.
- SDXL CLIP Encoder-1
- SDXL CLIP Encoder-2
- SDXL base 1.0_0.9vae
- LLaVA CLIP
- LLaVA v1.5 13B
- (optional) Juggernaut-XL_v9_RunDiffusionPhoto_v2
- Replacement of
SDXL base 1.0_0.9vaefor Photo Realistic
- Replacement of
- (optional) Juggernaut_RunDiffusionPhoto2_Lightning_4Steps
- Distilling model used in
SUPIR_v0_Juggernautv9_lightning.yaml
- Distilling model used in
-
SUPIR-v0Q: Baidu Netdisk, Google DriveDefault training settings with paper. High generalization and high image quality in most cases.
-
SUPIR-v0F: Baidu Netdisk, Google DriveTraining with light degradation settings. Stage1 encoder of
SUPIR-v0Fremains more details when facing light degradations.
- Edit Custom Path for Checkpoints
* [CKPT_PTH.py] --> LLAVA_CLIP_PATH, LLAVA_MODEL_PATH, SDXL_CLIP1_PATH, SDXL_CLIP2_CACHE_DIR * [options/SUPIR_v0.yaml] --> SDXL_CKPT, SUPIR_CKPT_Q, SUPIR_CKPT_F
RealPhoto60: Baidu Netdisk, Google Drive
Usage:
-- python test.py [options]
-- python gradio_demo.py [interactive options]
--img_dir Input folder.
--save_dir Output folder.
--upscale Upsampling ratio of given inputs. Default: 1
--SUPIR_sign Model selection. Default: 'Q'; Options: ['F', 'Q']
--seed Random seed. Default: 1234
--min_size Minimum resolution of output images. Default: 1024
--edm_steps Numb of steps for EDM Sampling Scheduler. Default: 50
--s_stage1 Control Strength of Stage1. Default: -1 (negative means invalid)
--s_churn Original hy-param of EDM. Default: 5
--s_noise Original hy-param of EDM. Default: 1.003
--s_cfg Classifier-free guidance scale for prompts. Default: 7.5
--s_stage2 Control Strength of Stage2. Default: 1.0
--num_samples Number of samples for each input. Default: 1
--a_prompt Additive positive prompt for all inputs.
Default: '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 Fixed negative prompt for all inputs.
Default: 'painting, oil painting, illustration, drawing, art, sketch, oil painting,
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 Color Fixing Type. Default: 'Wavelet'; Options: ['None', 'AdaIn', 'Wavelet']
--linear_CFG Linearly (with sigma) increase CFG from 'spt_linear_CFG' to s_cfg. Default: False
--linear_s_stage2 Linearly (with sigma) increase s_stage2 from 'spt_linear_s_stage2' to s_stage2. Default: False
--spt_linear_CFG Start point of linearly increasing CFG. Default: 1.0
--spt_linear_s_stage2 Start point of linearly increasing s_stage2. Default: 0.0
--ae_dtype Inference data type of AutoEncoder. Default: 'bf16'; Options: ['fp32', 'bf16']
--diff_dtype Inference data type of Diffusion. Default: 'fp16'; Options: ['fp32', 'fp16', 'bf16']# Seek for best quality for most cases
CUDA_VISIBLE_DEVICES=0,1 python test.py --img_dir '/opt/data/private/LV_Dataset/DiffGLV-Test-All/RealPhoto60/LQ' --save_dir ./results-Q --SUPIR_sign Q --upscale 2
# for light degradation and high fidelity
CUDA_VISIBLE_DEVICES=0,1 python test.py --img_dir '/opt/data/private/LV_Dataset/DiffGLV-Test-All/RealPhoto60/LQ' --save_dir ./results-F --SUPIR_sign F --upscale 2 --s_cfg 4.0 --linear_CFGCUDA_VISIBLE_DEVICES=0,1 python gradio_demo.py --ip 0.0.0.0 --port 6688 --use_image_slider --log_history
# Juggernaut_RunDiffusionPhoto2_Lightning_4Steps and DPM M2 SDE Karras for fast sampling
CUDA_VISIBLE_DEVICES=0,1 python gradio_demo.py --ip 0.0.0.0 --port 6688 --use_image_slider --log_history --opt options/SUPIR_v0_Juggernautv9_lightning.yaml
# less VRAM & slower (12G for Diffusion, 16G for LLaVA)
CUDA_VISIBLE_DEVICES=0,1 python gradio_demo.py --ip 0.0.0.0 --port 6688 --use_image_slider --log_history --loading_half_params --use_tile_vae --load_8bit_llava
@misc{yu2024scaling,
title={Scaling Up to Excellence: Practicing Model Scaling for Photo-Realistic Image Restoration In the Wild},
author={Fanghua Yu and Jinjin Gu and Zheyuan Li and Jinfan Hu and Xiangtao Kong and Xintao Wang and Jingwen He and Yu Qiao and Chao Dong},
year={2024},
eprint={2401.13627},
archivePrefix={arXiv},
primaryClass={cs.CV}
}
If you have any question, please email [email protected].
Replicate demo is based on https://github.com/chenxwh/SUPIR