chenydong / PA-Diff

The code of paper "Learning A Physical-aware Diffusion Model Based on Transformer for Underwater Image Enhancement"
Apache License 2.0
22 stars 0 forks source link

about method #4

Open xianfeizhu opened 5 months ago

xianfeizhu commented 5 months ago

Hello, In the provided code, paired SR and HR images are required as input for the physics prior generation branch module, both during the training and prediction phases. However, in practical applications, there are no paired SR and HR images available. The purpose of underwater image restoration is to obtain better HR images from SR images. In your code, you input HR images into the network for learning. The network only needs to learn to output the HR images you input to achieve the best restoration effect. Is there a potential issue with this approach?

ChenzhaoNju commented 5 months ago

Hello, In the provided code, paired SR and HR images are required as input for the physics prior generation branch module, both during the training and prediction phases. However, in practical applications, there are no paired SR and HR images available. The purpose of underwater image restoration is to obtain better HR images from SR images. In your code, you input HR images into the network for learning. The network only needs to learn to output the HR images you input to achieve the best restoration effect. Is there a potential issue with this approach? 这是代码写作的问题,实际上物理先验生成分支中GT只参与损失函数的构建,并不经过网络,因此不存在这样的问题。等有时间我们会修改代码,上传正确的版本。