Adaptive-step samplers (SA-Solver stochastic strength, predictor-corrector error estimate)
Nodes: Adaptive SA Sampler (exposes the stochastic strength and noise window that ComfyUI hardcodes in sa_solver), Adaptive PC Sampler (Adams predictor-corrector with a free local error estimate) and Trajectory Probe (per-step diagnostics). Research code with measurements.
Adaptive Samplers
ComfyUI samplers built around one idea: in a multistep predictor–corrector scheme the gap between the predicted and the corrected state is a local truncation error estimate that costs zero extra network evaluations, because the evaluation that produces the correction is the one the next step needs anyway.
This package exposes that estimate, and — more usefully in practice — exposes the stochastic strength of SA-Solver, which ComfyUI hardcodes.
Русский · Марий йылме · 中文
Install
Clone into ComfyUI/custom_nodes/ and restart ComfyUI. No extra dependencies; it uses torch,
numpy and safetensors, all already required by ComfyUI.
Nodes
| Node | Output | Purpose |
| --- | --- | --- |
| Adaptive SA Sampler | SAMPLER | SA-Solver with exposed stochastic strength (eta) and noise window |
| Adaptive PC Sampler | SAMPLER | Deterministic Adams predictor–corrector with free error estimate |
| Trajectory Probe | SAMPLER | Wraps any sampler and dumps per-step diagnostics to output/exp |
Sampler names are also registered for the plain KSampler dropdown and for API workflows:
asa_fixed, asa_adaptive, apc_2, apc_3, apc_2_pred, apc_2_fixed, apc_3_fixed.
The useful part: stochastic strength
ComfyUI ships SA-Solver (Xue et al., NeurIPS 2023) as sa_solver. It applies a multistep Adams
predictor–corrector to the drift only, keeps the diffusion term separate, and costs exactly
1 NFE per step. ComfyUI calls it with tau_func=None, which pins the stochastic strength to
eta = 1.0 inside a fixed 20–80 % window. There is no node to change that.
Adaptive SA Sampler exposes it. At eta = 1.0, response = 0.0 and window 0.2–0.8 it
reproduces sa_solver bit-for-bit (verified: maximum absolute difference 0.000e+00).
Measured behaviour
Anima (flow-matching DiT), 768×1024, cfg 5.0, normal scheduler, 10 NFE, 4 prompts × 2 seeds.
Geometry is the RMS deviation from a 100-step euler reference after 8× average pooling;
texture is the RMS of the high-pass residual. A converged 30-step SDE scores 0.3335 on texture.
| Sampler | Geometry (lower is better) | Texture |
| --- | --- | --- |
| euler (ODE) | 0.3677 | 0.1368 |
| euler_ancestral (SDE, no drift corrector) | 0.4016 | 0.1318 |
| sa_solver (SDE + drift corrector) | 0.3516 | 0.2244 |
The drift corrector gives a stochastic sampler better macro geometry than the deterministic solver at the same budget, while keeping 64 % more micro-relief.
Choosing eta
| eta | Geometry | Texture | Texture gap closed |
| --- | --- | --- | --- |
| 1.0 | 0.3516 | 0.2244 | 0 % |
| 1.4 | 0.3865 | 0.2503 | 24 % |
| 1.8 | 0.4225 | 0.2885 | 59 % |
| 2.0 | 0.4541 | 0.2975 | 67 % |
Texture bought per unit of geometry lost: 0.74 from 1.0→1.4, 1.06 from 1.4→1.8, 0.28 from
1.8→2.0. The exchange rate collapses after 1.8, and fine structures (eyelashes, leaf edges) start
turning into scribble. Useful range is 1.6–1.8.
Raising eta also shifts composition, not only surface detail — expect a different image, not the
same image with more texture.
Do not widen the noise window
Moving stochastic_start/stochastic_end from 0.2–0.8 to 0.1–0.9 costs about 20 % of macro
geometry and yields no texture at all (0.2244 → 0.2225). Widening further to 0.05–0.95
changes nothing. Noise injected at very high sigma is integrated away before it becomes structure;
noise injected at very low sigma has no steps left to resolve. Only the middle band converts
stochasticity into detail.
Trajectory Probe
Wraps any SAMPLER and writes a .safetensors per run into output/exp, containing per-step
sigma, pc_error (the free local error estimate), curvature, state/denoised/derivative norms,
the schedule, the final latent, and optionally a stack of intermediate latents (latent_stride).
The generating prompt graph is stored in the file metadata.
analyze_dumps.py decomposes the variance of log(pc_error) into shares explained by step index,
by cfg, and by the individual sample.
What did not work
Reported because it was measured, not assumed.
- Adaptive step size.
apc_*inbudgetmode redistributes steps using the free error estimate. It beats a uniform grid by a wide margin but loses to ComfyUI'snormalscheduler (0.81–0.99× its accuracy). Schedules derived offline by equidistributing the measured error also lose tonormal. Kept for ablation, not recommended for production. - Error-driven noise control (
response≠ 0). On every test it matched, within noise, simply loweringetaby the same average amount. Keepresponse = 0.0. - Solver order at production cfg. At cfg 1.0 a 4th-order method beats euler by 1.30–1.52×. At
cfg 5.0 the best of
dpmpp_2m/ipndm/deisbeats euler by only 1.01–1.04×, and at cfg 8.0 not at all. Truncation error is not the dominant error term at usual guidance scales.
Notes on other samplers
Measured on the same model, worth knowing before building on them:
uni_pcis hardcoded to a VP/eps parametrisation (noise / sqrt(1 + sigma²),model_type = "noise"). It is not valid for flow-matching models and scores 6.86 against 0.63 for euler at cfg 8. Do not use it with Flux/SD3/Anima-class models.- NFE per step, measured:
sa_solver,euler_ancestral,res_multistep_ancestral,dpmpp_3m_sde,er_sde,deis,clyb_4m_sde_momentumized— 1.0.heun,seeds_2,dpmpp_sde— 1.9.res_3s_ode— 3.0,res_5s_ode— 5.0,res_6s_ode— 6.0.dpm_adaptive— 3.0. Only 1-NFE samplers are comparable at an 8–14 NFE budget. - At 8 NFE, high-order stochastic multistep without a corrector destabilises badly:
res_multistep_ancestralscores 1.93 andclyb_4m_sde_momentumized7.74 on geometry, against 0.40 forsa_solver.
Tests
Standalone, no GPU required:
python test_pc_core.py # Adams weights, convergence orders, error calibration
python test_adaptive_step.py # adaptive vs static schedules on a nonlinear benchmark
Requires ComfyUI's environment, run from the ComfyUI root:
venv/Scripts/python.exe custom_nodes/adaptive-samplers/test_comfy_path.py # NFE accounting
venv/Scripts/python.exe custom_nodes/adaptive-samplers/test_stochastic.py # sa_solver parity
test_pc_core.py verifies that the Adams predictor/corrector weights match textbook coefficients
to 1e-16, that local truncation orders are 2/3/4/5 for orders 1–4, and that the free error estimate
tracks the true local error with a median ratio of 0.93–0.98.
Experiment scripts
run_grid.py, run_schedule_ab.py, run_sde_test.py, run_jitter_sweep.py drive ComfyUI over its
HTTP API using probe_variance.json as the graph template. They are the scripts that produced the
numbers above; adapt the model, LoRA and VAE names in the template before use.
Citation
@software{savelev_yuzvynsky_2026_adaptivesamplers,
author = {Iakov Savelev and Yehor Yuzvynsky},
title = {Adaptive Step Samplers: Zero-Cost Truncation Error Estimation and Exposed Stochastic Control for ComfyUI},
year = {2026},
publisher = {Zenodo},
doi = {10.5281/zenodo.23529356},
url = {https://doi.org/10.5281/zenodo.23529356}
}
License
MIT License, see LICENSE.