benchmarks evals
Eight Guidance Tricks Couldn’t Consistently Beat the Baseline
A re-evaluation on two rectified-flow transformers found no training-free guidance method that improved every measured criterion.
Summary
A re-evaluation on two rectified-flow transformers found no training-free guidance method that improved every measured criterion.
The study retested eight classifier-free-guidance relatives with compositional-alignment metrics. APG recorded several nominal best scores, often within uncertainty, while attention perturbations showed isolated gains and more frequent degradation on one model. Plain CFG remained a competitive lower-cost baseline in this protocol.
Why it matters
A re-evaluation on two rectified-flow transformers found no training-free guidance method that improved every measured criterion.
Limits and context
No additional limitation was separately recorded.
Key claims
A re-evaluation on two rectified-flow transformers found no training-free guidance method that improved every measured criterion.
Evidence: source-2026-08-18-019
Sources
- arXiv preprint 2608.16786arXiv · primary research
Corrections
No corrections have been recorded for this story.