Reproduce “SCALE: Improving Guidance-Free Visual Generation via Self-Contrastive Alignment for Likelihood Estimation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jiwan Hur, Dong-Jae Lee, Dongyeun Lee, Gyojin Han, Junmo Kim
SCALE reduces sampling-time guidance costs by training one model to produce a contrast-enhanced prediction. It uses differences between predictions on real and self-generated inputs; conditional generation incorporates classifier-free guidance during training. On ImageNet 256 × 256 with DiT-XL/2, SCALE reports FID-50k 1.87 versus GFT’s 1.99, using 50 evaluations rather than CFG’s 100; this is an evaluation count, not elapsed time.
Key insights
- The difference between a model’s predictions on real and self-generated inputs can serve as a signal for steering outputs toward regions better supported by the data.
- The framework applies to both diffusion noise prediction and autoregressive next-token prediction, and can combine self-contrastive enhancement with classifier-free guidance.
- Stronger enhancement does not improve every metric: in text-to-image tests, preference measures rose at lower pseudo-temperature settings while FID worsened.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.