wyhuai / DDNM

[ICLR 2023 Oral] Zero-Shot Image Restoration Using Denoising Diffusion Null-Space Model
MIT License
1.08k stars 75 forks source link

使用自己的数据(模糊的数据,不是imagenet,也不是人脸数据),何如使用作者论文提出的工作恢复图像质量 #48

Open shengshneg123 opened 12 months ago

shengshneg123 commented 12 months ago

作者工作十分优秀,在代码复现的过程中,有几个问题请作者解答。1.hq_demo中的推理和其文件夹之外的推理使用不同的权重不同的配置文件,这是为什么?2.模糊的数据,不是imagenet,也不是人脸数据,应该如何恢复,还需要在guided_diffusion 上训练权重吗?

shengshneg123 commented 12 months ago

期待作者的恢复》》》

wyhuai commented 12 months ago

Thanks, for question 1, In hq_demo, since repaint uses better pre-trained models, we used the code base of repaint to implement DDNM, so the config file is inherited from reapint. 2, In this case, it is better to train a new model, otherwise, the restoration quality may not be desirable.