CoFlow adds contrastive trajectory repulsion to Flow Matching, improving few-step generation on ImageNet
burny_tech · x · 2026-10-04
A new arXiv paper introduces CoFlow, which brings contrastive learning into Flow Matching by injecting a repulsive drift term during training that pushes trajectories away from negative ones, explicitly lowering local Lipschitz constants of the velocity field.
- Prior fixes treated trajectory crossings indirectly via post-hoc distillation or endpoint coupling; this paper links crossings to high-frequency velocity signals (hard to fit due to spectral bias) and severe numerical integration errors in few-step inference
- Formulated from an SDE perspective, with an equivalent stochastic interpolant formulation giving a tractable design space for controlling negative-sample influence
- Consistently boosts FM, REPA, and OT-CFM baselines on ImageNet for few-step generation
More from Research
- Protein watermarks survive scrutiny: researchers say synthesis providers can incentivize keeping them — anshulkundaje · 2026-10-04
- Bab, a BLAKE3-Inspired Hash Function Family With Streaming Verification, Goes Open Source — carsonfarmer · 2026-10-04
- From Text Tokens to Pixels: How Vision Encoders Turn Images Into Meaning — _jaydeepkarale · 2026-10-04
- Paper at NeurIPS: individual parameters in weight-sparse transformers appear interpretable — CatAstro_Piyush · 2026-10-04
- How vector databases work: embeddings, cosine similarity, HNSW, IVF and PQ explained — blaizedsouza · 2026-10-04
- ARROW unifies 3D reconstruction and point tracking from arbitrary image sets, sets new SOTA — CSProfKGD · 2026-10-04