[Resource] We unlocked SA-Solver in ComfyUI: why stochastic strength eta=1.7 beats Euler at 10 steps (and what didn't work)
retained text
Hey everyone, If you generate images on modern diffusion and flow-matching models (Flux, SD3, Anima) at low step counts (8–14 steps), you're probably stuck between two bad options: * **Euler:** Good macro anatomy, but surfaces look waxy and plastic. * **Ancestral samplers (**`euler_a`**, etc.):** Better texture, but anatomy drifts and degrades. ComfyUI actually has a brilliant solver under the hood called `sa_solver` (Xue et al., NeurIPS 2023). It runs an Adams predictor-corrector on the *drift* while keeping diffusion separate, costing **exactly 1 NFE per step**. But there was a catch: **ComfyUI hardcoded its stochastic strength (**`eta`**) to 1.0 inside a fixed 20%–80% window**, with no settings to adjust it. Together with my co-author Yehor Yuzvynsky (who provided compute hardware and editorial review), we created **Adaptive Step Samplers**: 👉 **GitHub:** [https://github.com/YakovSava/adaptive-step-samplers](https://github.com/YakovSava/adaptive-step-samplers) *(No dependencies required — uses stock PyTorch, NumPy, and SafeTensors).*
document fetched 2026-10-22 00:34 · canonical https://reddit.com/r/StableDiffusion/comments/1xbi156/resource_we_unlocked_sasolver_in_comfyui_why