← Reference notes

Global Transport Couplings for Classifier-Free Guided Flows

Summary

Global Transport (GT): pair noise and data with one class-agnostic minibatch optimal-transport coupling (labels ignored) when training conditional flow-matching models. It hurts unguided generation but, combined with classifier-free guidance, consistently improves quality and lets guidance scale go higher before degrading — no change to architecture, sampler or guidance (Abstract; §3–4).

When to apply

• Training a class- or text-conditioned flow-matching image model that will be sampled with CFG.
• Evaluate couplings under the guided sampler you'll actually use, not unguided (authors' main lesson).

Procedure

1. In each training minibatch, compute the Euclidean cost between noise samples and data samples (ignore labels).
2. Solve the assignment (exact OT) and train the flow on the re-paired (noise, data, condition) triples.
3. At inference, sweep CFG scale; expect optimum at higher w than with independent couplings.

Costs and limits

• Theory bound doesn't control the learned fields along CFG trajectories or guarantee a quality ordering; transfer outside image/video/cell domains open (Limitations).
• Worse than baseline without guidance.

Evidence (detail)

• ImageNet-256, SiT-XL/2: FID 2.06 (independent coupling) → 1.91 with GT (§4, Table 1).
• Gains across SiT B/2, L/2, XL/2, various NFEs, text-conditioned ImageNet-256, and single-cell gene expression (§4, Tables 2–3, Fig. 5).
• Overhead: O(B²) memory, O(B³) time per batch, usually <2% of training time but expensive for very large batches (§3).

Facts

• id: F1 claim: ImageNet-256, SiT-XL/2: FID 2.06 (independent coupling) → 1.91 with GT (§4, Table 1). confidence: medium # single source: author-reported date: 2026-10-06 sources: [S1] license: CC BY 4.0
• id: F2 claim: Gains across SiT B/2, L/2, XL/2, various NFEs, text-conditioned ImageNet-256, and single-cell gene expression (§4, Tables 2–3, Fig. 5). confidence: medium # single source: author-reported date: 2026-10-06 sources: [S1] license: CC BY 4.0
• id: F3 claim: Overhead: O(B²) memory, O(B³) time per batch, usually <2% of training time but expensive for very large batches (§3). confidence: medium # single source: author-reported date: 2026-10-06 sources: [S1] license: CC BY 4.0

Related

• OT-CFM / minibatch OT flow matching (Tong et al.). Not agent-related.

Review notes

• Built from private assessment notes; social-media hype check and source-post links omitted. Paraphrase only; no verbatim quotes beyond short attributed terms.
• Figures not re-checked against the PDF in this pass beyond the original private notes.

For language models and agents: start at /llms.txt. This page in Markdown: /library/reference/arxiv/2610-07555-global-transport-cfg-flows.md.