Reproduce “FACM: Flow-Anchored Consistency Models” 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
Yansong Peng, Kai Zhu, Yu Liu, Pingyu Wu, Hebei Li, Xiaoyan Sun, Feng Wu
FACM stabilizes continuous-time consistency training by anchoring a shortcut model to the instantaneous velocity field learned through Flow Matching. The method reaches FID 1.32 on class-conditional ImageNet at two function evaluations, while its Chain-JVP implementation enables consistency training of a 14-billion-parameter model within roughly 72 GB of GPU memory.
Key insights
- Explicit velocity supervision prevents the shortcut model from drifting away from the generative flow; removing the Flow Matching term causes reported FIDs of roughly 372–392 during early training.
- Separating Flow Matching and consistency conditioning through an expanded time domain is more effective than an auxiliary timestamp, reaching 1.81 versus 1.97 FID at one evaluation in the cited ablation.
- FACM primarily compresses a learned flow trajectory: stronger teachers improve the distilled model, while the method retains the original backbone rather than requiring specialized normalization or architectural changes.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.