Global Transport coupling lifts guided flow generation: ImageNet FID 2.06 to 1.91
mmbronstein · x · 2026-10-15
- A preprint from Oxford, Mila, MIT, AITHYRA and others (incl. Michael Bronstein) introduces Global Transport (GT), a class-agnostic optimal-transport coupling for flow matching models.
- Instead of random noise-image pairing, GT matches noise to images across the whole batch without class labels, reducing the gap between conditional and unconditional predictions under strong classifier-free guidance (CFG).
- Counterintuitively, GT alone worsens unguided generation, but combined with CFG it consistently improves quality across domains, model scales, and sampling budgets — suggesting couplings should be evaluated under the guided flow used at inference.
- On ImageNet-256, FID improves from 2.06 to 1.91 with no changes to architecture, samplers, or guidance mechanisms, offering a new training-time lever for conditional generation.
Related event: Global Transport Couplings Cut ImageNet FID to 1.91(3 posts)→
More from Research
- Watching AI evaluation shift from offline benchmarks to online eval — vboykis · 2026-10-15
- StartLux-Decision debuts RL-trained calibrated decision models, 27B scores 72.8 — startlux-models · 2026-10-15
- A third of all math preprints in 9 months were written with AI, researcher says — IgorCarron · 2026-10-15
- AAAI/ACM AIES paper interviews 15 data scientists on why responsible AI practice is hard to sustain — ChinasaTOkolo · 2026-10-15
- Schmidt Sciences awards $10M to 19 AI humanities research teams — OdedRechavi · 2026-10-15
- Reddit debate: is overfitting killing LLM creativity? — Expensive-Paint-9490 · 2026-10-15