Unlocking SA-Solver in ComfyUI: eta=1.7 beats Euler at 10 steps for low-step image generation
YakovSava · reddit · 2026-10-21
For low-step (8–14) generation on Flux, SD3, and Anima, users are stuck between Euler (good anatomy, waxy surfaces) and ancestral samplers (better texture, drifting anatomy).
The answer is ComfyUI's built-in sasolver (Xue et al., NeurIPS 2023): an Adams predictor-corrector on the drift with diffusion kept separate, costing exactly 1 NFE per step. The catch: ComfyUI hardcoded the stochastic strength eta to 1.0 in a fixed 20%–80% window with no exposed settings.
The author and co-author released Adaptive Step Samplers, finding eta=1.7 beats Euler at 10 steps and documenting what didn't work. No dependencies beyond stock PyTorch, NumPy, and SafeTensors.
Related event: SA-Solver in ComfyUI Beats Euler at Low Step Counts(2 posts)→
More from Multimodal
- MiniMax H3 r2v Tip: Skip Speed LoRA and Use at Least 20 Steps with Video Reference — theshield99 · 2026-10-21
- UMELINA, Bratislava's first AI film festival, opens free submissions Nov 2026 — lmoroney · 2026-10-21
- AI agent scores Steve Jobs' Stanford speech in GarageBand fully via computer use — itsOmSarraf_ · 2026-10-21
- EasyMOSS workflow makes bbox region control trivial on MOSS in ComfyUI — Quiet_Quit_8894 · 2026-10-21
- Fish Audio raises $52M and launches open-weight emotive TTS model Drama 3 — fekdaoui · 2026-10-21
- Claude Code renders lyric videos frame-by-frame in Chromium, no AI video model needed — aitrendz_xyz · 2026-10-21