Question about perceptual model usage and the impact of perceptual/GAN losses

[chenxx89]

First of all, thank you for sharing your impressive work and making the code publicly available.

In the training code, I noticed that there seems to be an attempt to utilize the DINOv3 model for calculating perceptual loss; however, it appears that this component is not ultimately activated or used in the final implementation. Could you kindly share the reasoning behind this decision? Specifically, was it due to DINOv3 yielding inferior results compared to LPIPS in your experiments?

In my previous experiments about restoration tasks based on one step diffusion , I have observed that perceptual loss plays a significant role in enhancing image clarity. Therefore, I am very interested in understanding how perceptual loss and GAN loss respectively contribute to the overall performance in your framework. Any insights you could provide regarding their relative impact on the final results would be greatly appreciated.

Thank you very much for your time and guidance.

[double8fun]

We have not conducted an in-depth study of dinov3 as perceptual loss, so we are unable to draw any definitive conclusions on this point.
As for the respective contributions of perceptual loss and GAN loss, this is an interesting topic. Based on our observations, using only perceptual loss on a single-step DiT yields good degradation removal results but leads to grid-like artifacts and a loss of generative capability. Introducing GAN loss is a better solution.