Rebuilding the model · image formation
Evolving denoiser
Build a denoiser that can say which detail belongs to the scene, which belongs to corruption, and why.
Current question
What joint law can carry radiance, residual, scale, and causal ancestry without silently deleting real detail?
What exists
The work has already killed two attractive answers: the original FMMT denoising model and a patch-typicality branch. Both remain documented because a useful research program remembers why its discarded mechanisms failed.
The active replacement decomposes the entire observation before removing its fine component. It tracks radiance, residual, and causal ancestry across scale. The decomposition order survived; the first confidence recurrence did not. No denoiser is being passed off as finished.