paper:
https://arxiv.org/pdf/2308.15070
codes:
https://github.com/xpixelgroup/diffbir
Motivation
- diffusion生成先验在条件图像生成效果很好
- 直接用LQ做条件输入然后生成的效果不好,因为LQ损伤严重,特征纠缠,会造成伪像残留
- 将任务分解为两部份,1 退化去除 2 特征生成
- 退化任务与退化设计有关,对于不同的任务,应用不同的退化去除模块,只做去退化,不做生成,MSE loss 训练
Main
- Blind real world degradation
- Unified-framework for deblur/sr/dehaze/de-rain/inpaint
- Two stage: degradation removal and details generation
- IRControlNet leverage the un-realistic details
- Region-adaptive module in infer process to guide denoising
只对低频域进行MSE loss约束,高频细节进行生成 - Different task different degradation removal module
Network
IRControlNet
Region-adaptive restoration guidance
Relative Tech
- SD
- VAE
- ControlNet
Eexperiment config
Paper Results and Performance
- DR2
- DiffFace
- PGDiff