Label-free batch-level optimal transport couplings stabilize strong CFG, FID 2.06→1.91
AlexanderTong7 · x · 2026-10-14
A paper highlighted by AlphaXIV, "Global Transport Couplings for Classifier-Free Guided Flows," tackles a core issue in flow matching: random noise-image pairings produce unstable trajectories under strong classifier-free guidance (CFG). The work instead uses optimal transport to match noise with images across the entire batch, ignoring class labels.
This shrinks the gap between conditional and unconditional predictions, enabling stronger guidance without quality loss. On ImageNet-256 it improves FID from 2.06 to 1.91 with no architecture or sampling changes. The authors also found these simple label-free mini-batch OT couplings improve conditional generation at higher guidance scales.
Related event: Global Transport Couplings Improve Classifier-Free Guided Flows at High CFG(2 posts)→
More from Research
- Qdrant tests Perplexity's pplx-embed-v2-late: 21x smaller vectors, 99.9% retrieval quality kept — qdrant_engine · 2026-10-14
- Always Look at the Data: Run Dumb Baselines to Catch Eval Set Biases — antoine_chaffin · 2026-10-14
- 0-Parameter 'Pick the Longest Answer' Heuristic Tops HLE Decision Benchmark at 20.7% — antoine_chaffin · 2026-10-14
- A 0-parameter 'pick the longest option' model tops HLE, exposing benchmark flaw — antoine_chaffin · 2026-10-14
- Gemma 4 E4B + genomic model BOTANIC-1 pinpoints causal melon yield mutation in one afternoon — remilouf · 2026-10-14
- ETH/EPFL release 134-page fully open Apertus v1.5 multimodal reasoning report — LucaAmb · 2026-10-14