ablation experiments for latent-replacing pretraining strategy

[Sttrr]

I don't see the ablation experiments in the paper “Refusion: Enabling Large-Size Realistic Image Restoration with Latent-Space Diffusion Models” about the latent-replacing pretraining strategy and I would like to know how much impact this strategy brings， can you provide comparison results?

[Algolzw]

Hi, we didn't do the ablation experiment for the "latent-replacing strategy". But intuitively, only in this way, we can combine LQ hidden connections with the restored HQ latent to generate the HQ image. Then we can perform diffusion-based restoration on the latent vectors.