Comparison with Directly Fine-tuning on LRP-processed Images

[liuyu19970607]

Hello team,

This is truly excellent and inspiring work. Congratulations on the great results!
I'm curious about the advantages of this specific design choice compared to an alternative strategy: directly fine-tuning the Kontext model using the LRP-processed images themselves as the training input.

It seems that the latter approach would also provide the model with the desired "attribution" signal. I'm wondering if there's a specific reason (e.g., better performance, training stability, efficiency, or generalization) why the proposed method was chosen over this direct fine-tuning approach.

Any insights you could provide would be greatly appreciated. I'm very much looking forward to your response.

Thank you!

[FeiSong123]

Thank you for your interest.

The main reason we chose to input both the low-quality image and the LRP-processed image together is based on our experimental observations. We found that using either input alone led to suboptimal results: using only the low-quality image often resulted in incomplete degradation removal, while using only the LRP image tended to produce over-smoothed outputs lacking in detail. Therefore, we adopted the strategy of simultaneous input to combine the strengths of both.

In addition, we also incorporated modulation based on timestep and the layer index in DiT. For more detailed information, you can also refer to our report.

Best regards