Does the task of diffusion-based dense prediction come to an end?

[Boatsure]

Thanks for your comprehensive ablation experiments！

Since Marigold proposed the terminology “monocular depth estimation based on the diffusion model”, the research in this sub-field has been advancing rapidly. One article after another is brilliant, but they all seem to reveal one fact: that based on the standard diffusion process, the less it resembles the original diffusion and the closer it is to the discriminative method, the higher the accuracy of the depth or normal estimation.

When the forward process becomes single-step and does not require text prompt, and the denoising process becomes single-step and does not require the participation of noise, does GenPercept still belong to diffusion methods? Has the ship of Theseus entered the ocean of discriminative method?

Specifically, at what level does the GenPercept belong to the diffusion model?