A lattice method predicts better decoding schedules before generation
In plain words
A model can generate text or images by revealing several positions at once. Which positions it reveals together can affect quality, even when its weights and number of steps stay the same.
This paper introduces a method that estimates how strongly positions depend on one another, then uses that estimate to compare reveal schedules before generation.
On ImageNet images generated in eight steps, the method ranked three reveal orders the same way as the model’s image-quality test: spread and random beat raster order. That supports this comparison for the tested image model, not every model or task.
The research
If a model has a fixed number of generation steps, which positions should it update together? Existing designs can leave that choice fixed, while quality is known only after generation.
A way to compare schedules beforehand could help explain why the same model performs differently under different reveal orders.
The method represents decoding schedules as paths and ranks them by estimated dependence discarded when positions update in parallel.
From pretrained weights, the method estimates pairwise dependence between positions and sums it along a schedule. This lower bound on dependence lost by parallel updates ranks schedules before generation.
On ImageNet-256, at eight steps, the released MAR-B evaluator’s image-quality score (FID; lower is better) was 139.80 for raster, 13.02 for random, and 9.44 for spread order. The estimate ranked them the same, supporting comparison in this setting.
The cost measures dependence discarded by parallel updates, but it leaves out errors in the model’s trained conditional predictions. The pairwise estimate is exact only when the stated first-order Markov condition holds, and some tested MAR-B rankings disagreed with its estimates.
Analysis
Deep
- Authors
- T. Y. Tsui, Jiatao Gu, Lingjie Liu
- License
- arXiv non-exclusive