Image feedback helps noisy-target reconstruction, but not reliable hole filling

The question

Hide a rectangle in a photograph, ask the trained component to fill it by constructing VSA vectors, then decode the result. Compare two ways of training that component: vector error alone, or vector error plus feedback from the decoded pixels.

Why this experiment?

Lower vector error did not guarantee better images. A matched training comparison asks whether feedback through the fixed decoder improves the actual missing-region task.

How it works

Hide a rectangle in a photograph, construct VSA vectors to fill it and decode the result. Fit two otherwise matched predictors, adding image-feedback loss to only one. Shuffling assigns encoded regions to wrong positions. The reduced-conditioning control removes photograph input from the predictor, while the decoder and preserved visible regions still use the photograph.

What happened

Adding image-reconstruction feedback did not qualify the trained component for hole filling. Mean hidden-interior error rose by 10.9%, but the paired uncertainty interval includes both improvement and harm. Mean error in already visible pixels rose by 30.8%, with 31 of 32 photographs worsening. In the separate diagnostic that supplies corrupted clean-target information, hidden-pixel error fell by 19.9% to 29.8% against the matched model across the three tested noise levels. That denoising improvement did not transfer to reliable completion from a masked photograph. Complete visual review found no reliable coherent filling. Independent audit passed; the recovered comparison is closed.

Limits and next question

This is one fixed loss intervention, not a search for the best architecture or coefficient. The codec saw the training collection. Fresh development and test sources are outside recorded prior exposures. Privileged noisy-target denoising is separate from target-denied completion. Numerical improvement alone cannot establish coherent completion; the 363-encounter life remains fixed. This is an already exposed panel from the interrupted comparison, not a fresh independent test or a second replicate. On preserved figures, 'No learned regional context' means 'No image context supplied to predictor': the same trained network receives zeros for both regional image vectors and its coarse image overview. Its learned weights, geometry, visibility mask and noise remain. After prediction, wholly visible regions are restored, and the final renderer still receives the shared masked overview. The photograph therefore remains context for the overall system. The causal question is: Does supplying the photograph to the predictor help beyond the photographic information already supplied to the decoder? This is neither an untrained model nor a final image produced without any context.

Next question. Interpret the fixed objective comparison before selecting another task. Video interpolation is a proposal, not an additional authorized or completed experiment.