CompFlow: Composing Velocity Fields for Multi-Condition Generation
Published in 2nd Workshop on Compositional Learning (CompLearn) at ICML 2026, Seoul, South Korea, 2026
A training-free flow-matching framework for compositional inference. Conditional velocity residuals are shown to be proportional to gradients of noised conditional likelihoods, so their weighted addition implements a tempered Product-of-Experts — generalising classifier-free guidance from one condition to arbitrarily many. On CLEVR, a model trained only with single-object conditioning composes shape, color, and position for up to five objects at 99.1–86.5% per-object accuracy, using 30× fewer network evaluations than prior baselines. Applied zero-shot on top of FLUX.1[dev], it improves the Share-CoT score on the T2I-CompBench non-spatial split from 0.7809 to 0.8021.
Project page: compflow.vornao.com
OpenReview: openreview.net/forum?id=JdQTzZewsy
Recommended citation: Miglior, L., Gervasi, V., & Bacciu, D. (2026). "CompFlow: Composing Velocity Fields for Multi-Condition Generation." 2nd Workshop on Compositional Learning at ICML 2026, Seoul, South Korea.
Download Paper