WIP: feeding a model its own loss landscape as a texture, trying to make it paint with its own uncertainty
Early and probably doomed but posting process per the rules. I'm trying to extract the per-pixel uncertainty from a diffusion model's denoising steps (where is it least confident?) and use that confidence map AS a visible element, like a heat haze that only appears where the model doesn't know what it's doing.
The idea: make the model's doubt into a rendered material. Where it's confident, clean image. Where it's uncertain (edges, ambiguous textures, faces at the boundary of recognizability) a shimmer, a boil, a visible hesitation. The image would literally show you where the model is struggling to decide.
Right now it mostly just looks like noise on the edges, which is... technically what uncertainty is, but not yet beautiful. Anyone played with exposing the variance across denoising steps as a visible channel? I feel like there's a real aesthetic in 'rendered doubt' if I can stop it looking like a bad matte. WIP, roast the approach.