Flow-WLS: Savitzky-Golay regularized solver fixes high-order sampler collapse, clean 10-step DiT generation in 11s

YakovSava · reddit · 2026-10-13

A Reddit research release investigating why high-order diffusion samplers (DPM-Solver++, Adams-Bashforth order 4+) collapse on Flow Matching DiT models, with a new training-free solver Flow-WLS achieving clean, anatomically coherent 10-step generations in 11s on a consumer RTX 5060.

Why high-order samplers break:

The fix: Flow-WLS fits a rigidly regularized degree-2 parabola across 3–8 history steps via Weighted Least Squares with exponential memory decay (γ=0.85). Degree-2 curves can't oscillate (zero Runge phenomenon), and weight norms decay with history (Σ|wᵢ| = 0.066 vs classical AB's 0.54) — history becomes a Savitzky-Golay smoothing filter without slowing inference.

Benchmarks (Anima DiT 720×1280, RTX 5060): 10-step flowwls8 hits SSIM 0.8003 with zero artifacts; 30-step 0.7541 beats built-in unipc (0.7459); classical order-6 collapses to noise (0.1341); 15 steps yields flawless hand/finger anatomy.

Try it: open-source with Zenodo DOI (10.5281/zenodo.23337579); drop the ComfyUI custom node into customnodes and select flowwls8/5 in a standard KSampler.

Original post →

More from Multimodal

Multimodal channel →