NAFNet
NAFNet copied to clipboard
megvii-research
The state-of-the-art image restoration model without nonlinear activation functions.
I found that NafNet and HINet use the conv with bias=True. But the MPRNet used the conv with bias=False. Since the HINet shares the same multi-stage architectures as MPRNet, I...
Hi, thanks for your excellent work! I am interested in Raw Image Denoising recently~ The code you provided does not include anything about Raw Image Denoising!!! Can you provide the...
Hello，您好！我有个问题是关于您所写论文里的结果对比部分，具体地说是Table 6. 为什么NAFNet在GoPro数据最终测试的结果由32.85变为了33.69呢？我想我应该认真阅读了您的整篇文章。是因为实验部分采用了与MPRNet-Local方法一致的TLSC方法吗？ 期待您的答复！
看了一下论文，我自己也训练了一遍width 32的网络，结果能和论文对上，但是貌似结果是SIDD validation set的值？是否不是benchmark的结果呢？ 期待您的回复！
Thank you for your wonderful work. I use four blocks of 3090，and batch size per gpu is 4. It cost about 15 hours training NAFNet-width32 on SIDD, and about 3...
关于loss疑问
疑问1: 在代码image_restoration_model.py 225行l_total = l_total + 0. * sum(p.sum() for p in self.net_g.parameters()) . 0乘以任何数都=0,这是啥操作? 疑问2: 关于PSNRLoss:self.scale * torch.log(((pred - target) ** 2).mean(dim=(1, 2, 3)) + 1e-8).mean(). 即 10/ln(10) *...
Metadata
Owner
Metadata
The state-of-the-art image restoration model without nonlinear activation functions.