Renormalization Group Flow Matching Scales Local Generative Modeling

Summary

Masuki and Ashida propose RGFM, a generative framework that employs the renormalization group as an exact probability path in flow matching. This approach enables local generation at O(ln L) patch sizes and near-linear scaling while provably preserving long-range correlations. Numerical results on 1B distributions and FFHQ images demonstrate significantly better consistency compared to standard local flow matching.

Why it matters

Generative modeling has been stuck between expensive global models and cheap local models that lose long-range structure. RGFM offers a physics-based, theoretically sound escape from this trade-off. If the O(ln L) locality guarantee holds broadly, high-fidelity generation could become feasible on modest hardware without sacrificing global consistency. The work also establishes a rare, rigorous bridge between RG theory in statistical mechanics and modern generative machine learning.

Highlights

• RGFM proves that its probability flow can be approximated by local velocity fields with a range of O(Λ⁻¹[ln L + ln(1/ε)]), enabling O(ln L) patches at near-linear cost.
• On FFHQ, RGFM produces significantly more consistent 64×64 and 256×256 samples compared to local flow matching, which exhibits large errors over long distances.
• The framework relies on an *exact* RG flow as the probability path, providing a principled multi-scale generation ordering from long to short wavelengths.

Critical take

Experimental validation is weak: only 1B toy distributions and FFHQ at modest resolutions were used. Since no comparison was made against the best global diffusion or large-scale flow matching baselines, the practical advantage over simply scaling a global model cannot be quantitatively determined. The entire structure depends on the existence of an exact RG flow that behaves well for the target distribution. This assumption is not guaranteed for arbitrary, high-dimensional, or discrete data, and the paper does not adequately test this prerequisite.