Reddit experiment: adaptive per-step noise in ComfyUI samplers barely helps, fixed schedule matches it

YakovSava · reddit · 2026-11-01

The author built a ComfyUI sampler node (Flow-SDE Sampler, a modified SA-Solver) that spends a fixed total noise budget across steps, and tested whether adaptively choosing per-step noise from the trajectory beats a constant eta at 6–15 steps on Flux 3 Klein, MOSS-Image and Anima, always against constant-eta runs with identical total noise.

Findings:

Curvature-based allocation does nothing reliable on any of the three models.
The solver's predictor–corrector error estimate gives a small gain (1–2% of a texture metric) on Flux 3 Klein, confirmed on fresh seeds; one-sided on MOSS-Image, undetectable on Anima; Krea 2 inconclusive (too slow).
A fixed schedule achieves the same gain: slightly more noise at the start and end of the window, less in the middle — no online adaptation needed.
In a 39-person human study, people could tell eta 1.4 vs 2.0 apart but could not distinguish the new allocation from plain noise.

Practical takeaway is modest: with sasolver, a bit more noise at window edges may slightly increase texture at equal composition drift. Metrics are latent-space proxies, not quality judgments. Node, data, per-image metrics and all negative results are on GitHub (with Zenodo DOI).

Original post →

Reddit experiment: adaptive per-step noise in ComfyUI samplers barely helps, fixed schedule matches it

More from Multimodal