Human vision has extreme spatial non-uniformity: high resolution at the fovea, low resolution in the periphery. Image generation models ignore this entirely, producing uniform resolution across the frame — most of which the viewer will never perceive at full detail.
Note (2026-11-10): I checked the abstract and searched the full text of arXiv 2603.23491. The paper reports speedups of up to 2x and describes the token count and generation time as drastically reduced; the essay's statement that the compute saving is proportional to the fraction of the image in the periphery is the essay's own, and the paper's setting is one where computational cost grows quadratically with the number of tokens. The paper says the foveated results are perceptually indistinguishable from full-resolution generation, that a naive mixed-resolution baseline shows structural artifacts at the region boundaries, and also that its own method shows occasional colour artifacts near the foveation boundary (from blending the decoded low- and high-resolution content). It also reports that FID may not be a reliable metric here, since the naive baseline scored better than full-resolution generation on it.
The authors (arXiv:2603.23491) exploit this with foveated diffusion. Eye-tracking data determines where the viewer is likely to look. A foveation mask guides token allocation: more tokens where attention concentrates, fewer in the periphery. The result is perceptually indistinguishable from full-resolution generation while using substantially fewer tokens and less computation.
The key technical challenge: mixed-resolution generation must not produce seams between high- and low-resolution regions. The authors develop a training method that creates mixed-resolution tokens from high-resolution data, enabling post-training of existing diffusion models.
The through-claim: generating more resolution than the viewer can perceive is pure waste. The uniform-resolution assumption isn't a physical requirement — it's a convenience of existing architectures. Once generation can be spatially adaptive, the compute savings are proportional to the fraction of the image that falls in the visual periphery, which for most content is most of the image.