Unlocking ComfyUI's SA-Solver: eta=1.7 beats Euler at 10 steps
YakovSava · reddit · 2026-10-21
YakovSava and Yehor Yuzvynsky released Adaptive Step Samplers (open source, stock PyTorch/NumPy/SafeTensors only), unlocking ComfyUI's built-in sasolver (Xue et al., NeurIPS 2023) whose stochastic strength was hardcoded to eta=1.0 in a fixed 20%–80% window. At low step counts (8–14), Euler gives waxy surfaces while ancestral samplers drift anatomically; sasolver runs an Adams predictor-corrector on the drift at exactly 1 NFE per step. Their finding: stochastic eta=1.7 beats Euler at 10 steps, and they document what didn't work.
Related event: SA-Solver in ComfyUI Beats Euler at Low Step Counts(2 posts)→
More from Multimodal
- Fibo2 old-photo restoration and colorization tested on a historic fire scene — Striking-Long-2960 · 2026-10-21
- Full Seedance 2.5 prompt shared for generating ultra-realistic handheld smartphone video — techhalla · 2026-10-21
- GPT Image 2.5 vs Nano Banana 2.1: 502 synthetic rust masks reviewed, 79.7% retained — Gloomy_Recognition_4 · 2026-10-21
- AI-generated 'Doner Kebab Taxi' video goes viral, poster holds the music rights — average1804 · 2026-10-21
- Fibo 2.0 ComfyUI workflows: Edit Turbo is fast and solid at image editing — nomadoor · 2026-10-21
- MOSS t2i anti-blur tip: write 'photography', not 'photograph' — Bulky_Chemistry_196 · 2026-10-21