Adaptive per-step noise barely helps: Flow-SDE sampler tests on Flux 3 Klein, MOSS-Image, Anima
YakovSava · reddit · 2026-11-01
The author built a ComfyUI Flow-SDE Sampler node (modified SA-Solver) and tested whether per-step noise chosen from trajectory curvature or the solver's predictor–corrector error beats constant eta at 6–15 steps, with total noise budget held equal. Findings: curvature does nothing reliable; error-based adaptation gains 1–2% on a texture metric only for Flux 3 Klein; a fixed schedule (more noise at window edges) matches it. A 39-person blind study couldn't tell the new allocation from plain noise. Node, data and full negative results are open-sourced with a DOI.
Related event: ComfyUI Test Finds Adaptive Noise Sampling Offers Little Gain(2 posts)→
More from Multimodal
- Dev shows off AI-generated game engine material graphs built from text descriptions — D3VAUX · 2026-11-01
- Seven sequential fine-tunes on Krea 2: stunning photorealism, catastrophic forgetting of everything else — MellyDArt · 2026-11-01
- TITAA #93: hands-on with Kling v4 and Mischief-1 video models from $0.01/sec — arnicas · 2026-11-01
- Comfy-Node-App: open-source extension builds custom app UIs for ComfyUI subgraphs — Dartium1 · 2026-11-01
- Google's StreamVibe: a video agent that watches you work — but no model times its help reliably — The Decoder · 2026-11-01
- AI short film made with Midjourney and Krea rakes in 200M views in 36 hours — LinusEkenstam · 2026-11-01