REVIEW 4 major objections 6 minor 10 references
FSampler tries to establish that a diffusion sampler can skip 15-25% of its model calls by extrapolating the denoising signal with low-order finite differences, cutting wall-clock time 8-22% with no retraining.
Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →
FSampler accelerates diffusion sampling by substituting finite-difference extrapolations of epsilon for model calls on selected steps, reducing NFE by 15-25% at SSIM 0.95-0.99.
T0 review reviewed 2026-08-03 challenge →
load-bearing objection Plausible engineering trick—epsilon extrapolation to skip model calls—but single-seed evidence and missing baseline comparisons leave the headline numbers unverified; send to review, not desk reject. the 4 major comments →
FSampler: Training Free Acceleration of Diffusion Sampling via Epsilon Extrapolation
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
Core claim
The paper's central claim is that the denoising signal epsilon varies smoothly enough along a sampling trajectory that the next model output can be predicted from the last few real outputs by finite-difference extrapolation. FSampler uses second-, third-, or fourth-order extrapolation to synthesize epsilon on selected steps, feeds the predicted value into the sampler's unchanged update rule, and thereby skips model calls. On three modern text-to-image models, representative configurations with a learning stabilizer reduce model calls by 15-25% and wall-clock time by 8-22% while keeping SSIM at 0.95-0.99; the adaptive gate can reach 45-50% fewer calls but only at SSIM around 0.73-0.74.
What carries the argument
The load-bearing object is the noise residual epsilon, defined as predicted clean image minus noisy latent. FSampler keeps a history of epsilons from real model calls and applies second-order (h2: 2*x[n-1]-x[n-2]), third-order Richardson (h3), or fourth-order (h4) finite-difference extrapolation, falling back to lower order when history is insufficient. The predicted epsilon replaces the model output on a skipped step while the sampler update is left untouched. Stability comes from validation checks (NaN and magnitude floors), a learning stabilizer that rescales predictions by an EMA ratio, and guard rails including protected head/tail windows, periodic anchors, and a cap on consecutive skip
Load-bearing premise
The load-bearing premise is that epsilon varies smoothly enough between real model calls that low-order finite-difference extrapolation in step index predicts the next value accurately; if the trajectory has sharp curvature, such as at a scheduler transition, the wrong predictor order or an unguarded skip will degrade fidelity.
What would settle it
Re-run the model at a step FSampler skipped and compare the true epsilon to epsilon_hat. If the L2 error between them is large on a schedule with smooth spacing while SSIM stays above 0.95, the claimed quality mechanism is not the extrapolation itself but something downstream; if even one unguarded skip in a protected-window-free run produces a visible artifact, the smoothness assumption fails on that schedule.
If this is right
- At conservative settings on the first tested model, a second-order pattern with the learning stabilizer reaches SSIM 0.9818 with 15.9% wall-clock time saved, while a more aggressive pattern reaches SSIM 0.9533 with 21.6% time saved.
- The method transfers to models with different samplers and schedulers; on a two-stage scheduler, third-order patterns handle the curvature discontinuity better than second-order.
- Because sampler update formulas are untouched, FSampler's savings are additive with distilled or quantized models and with any underlying scheduler choice.
- The adaptive gate marks a stability boundary: increasing acceptance to ~45-50% NFE reduction drops SSIM to ~0.73-0.74 on one model, so aggressive skipping trades visible fidelity.
- Higher-order predictors (h3, h4) do not consistently beat second-order on the Euler-like sampler, so more history does not automatically improve quality.
Where Pith is reading between the lines
- Because extrapolation happens in step index rather than in log-SNR time, the method should be stress-tested on schedules with irregular or data-dependent step sizes; the two-stage scheduler already shows that curvature breaks h2 but is handled by h3.
- The learning stabilizer's EMA ratio could double as a per-trajectory diagnostic: when the ratio pushes against its [0.5, 2.0] clamp, that signals a region where extrapolation is unreliable and real calls should be forced.
- A video-diffusion extension would be a natural test: temporal coherence might smooth epsilon further, but scene cuts could create the same curvature discontinuity seen in the two-stage scheduler, likely requiring per-scene predictor selection.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes FSampler, a training-free execution layer for diffusion sampling that reduces NFE by replacing some model calls with finite-difference extrapolations of the noise residual (epsilon). The method supports second-, third-, and fourth-order predictors, fixed or adaptive skip policies, validation checks, a learning stabilizer (EMA rescaling), and an optional gradient-estimation correction. Experiments on FLUX.1-dev, Qwen-Image, and Wan 2.2 report SSIM 0.95–0.99 with 15–25% NFE reduction and 8–22% time savings, and more aggressive skipping at SSIM ~0.73–0.74. The authors position the method as sampler-agnostic and orthogonal to other acceleration techniques.
Significance. If substantiated, FSampler would offer a simple, training-free, sampler-agnostic way to reduce diffusion sampling cost, with the significant practical advantage of dropping into existing ComfyUI workflows without modifying models or sampler formulas. The paper provides code and experiment data links, and the notation map is helpful for reproducibility. However, the current empirical evidence is too thin to support the headline fidelity and generalization claims: each configuration is run once on a single seed, the only perceptual metric is SSIM against a same-seed baseline, and the method's core extrapolation assumption is not directly validated. The contribution is potentially useful but requires substantially stronger evidence before the stated claims can be accepted.
major comments (4)
- [§3.1, §4.4] The extrapolation formulas (3.1)–(3.3) assume uniform spacing of 1 in step index, but the actual integration coordinate is log-SNR with nonuniform step sizes. The paper defines log_snr_step in §2 but never uses it in the predictors; for a two-stage scheduler (e.g., Wan 2.2's beta+bong_tangent), the coefficients are incorrect for nonuniform spacing. The paper's own admission of a 'curvature discontinuity' at the scheduler transition is exactly where this assumption is most fragile. The authors should either derive step-size-aware coefficients or provide direct predicted-vs-actual epsilon error measurements on skipped steps to justify the unit-index extrapolation.
- [§4.1, §4.2] Each configuration is executed once per model with a single seed. No confidence intervals, repeated trials, or multi-prompt evaluation are reported. Quality is measured by SSIM/RMSE/MAE against the same-seed baseline, which is a reproducibility measure, not a perceptual one; the abstract's claim that changes are 'often imperceptible on first glance' is unsupported without human or perceptual-metric evaluation. At minimum, the authors should run multiple seeds, report variance, and include a learned perceptual metric (e.g., LPIPS) or a small human study.
- [§3.3, §4.3] The validation procedure only rejects NaN/Inf and magnitude extremes; it does not measure directional accuracy, and the adaptive gate's h3-vs-h2 discrepancy is a heuristic, not a bound. The ablation in §4.3 shows that all four adaptive modes—learning, grad_est, learn+grad_est, none—produce identical SSIM=0.9533 on h2/s3, indicating that the stabilizers do not affect quality on FLUX.1-dev. This makes it unclear whether the reported fidelity comes from the extrapolation itself, the periodic anchors, or the protective windows. The paper should decompose the contribution of each component (including removing anchors/guard rails) and report the actual error of accepted epsilon_hat predictions on skipped steps.
- [§5 (Limitations)] The limitations section explicitly acknowledges single-seed comparisons and that SSIM does not capture all perceptual dimensions, yet the abstract and conclusion make strong generalization claims ('Across FLUX.1-dev, Qwen-Image, and Wan 2.2'). Additionally, the learning stabilizer beta is hand-tuned per model (β=0.9985 vs 0.995) and the best skip pattern differs per model (h2/s4 on FLUX, h2/s5 on Qwen, h3/s5 on Wan), so the 'sampler-agnostic' claim is overstated. The authors should present results for a single fixed default configuration (with no per-model hyperparameter tuning) and report sensitivity to beta, anchor_interval, and protected-window settings.
minor comments (6)
- [§3.2] Explicit skip indices: '0 and 1 are never skipped' should be clarified—is this because of the history requirement (min 2 real epsilons for h2) or a deliberate policy? Also explain interaction with protected windows and max_consecutive_skips.
- [§2] The variable 'time' is overloaded: it is used both as the ODE integration variable (dx/dtime) and as a step-size difference (sigma_next - sigma_current). Rename the step size to 'dt' or 'h' to avoid confusion.
- [§4.3] The statement 'gradient-estimation overhead does not translate into quality gains here' is vague; report the actual wall-clock times for learn+grad_est and learning modes so the reader can verify the overhead claim.
- [General] The paper lacks a related-work section. PFDiff and Timestep Embedding are cited in the bibliography but never discussed; add a comparison to these training-free acceleration methods and to cache-based approaches.
- [§4.1] The Qwen-Image baseline is described as '25 model calls across 50 schedule transitions'; clarify this (presumably two transitions per sampling step) and how this matches the NFE=25 definition.
- [§3.3] The learning_ratio update has no initial value specified; state the default (e.g., 1.0) and the exact clamping range (stated as [0.5,2.0]) in the algorithm description.
Circularity Check
No significant circularity: FSampler's central prediction is a fixed finite-difference extrapolation of stored real epsilons, and quality claims are measured against same-seed baselines rather than derived from fitted parameters.
full rationale
The paper's central prediction step (Section 3.1) is definitionally non-circular: epsilon_hat is a direct finite-difference extrapolation of previously computed real model outputs (h2/h3/h4), with no parameter fitted to the target output and no target quantity encoded in the predictor. Validation checks and stabilizers (Section 3.3) are guard rails that operate on the predicted epsilon or on differences between predictors; the learning ratio is an online EMA updated from real steps and applied only to subsequent skipped steps, so it does not make outputs equal to the baseline by construction. The paper's reported NFE reduction is a counting consequence of the chosen skip cadence, but it is presented as a design choice, not as a derived prediction; the load-bearing empirical claims are the measured SSIM/RMSE/MAE values against same-seed baselines, which are external to the method's equations. There are no self-citations, no imported uniqueness theorems, and no ansatz smuggled in via citation. The acknowledged limitations (single-seed comparisons, per-model selection of beta and cadence, SSIM as primary metric, and the Wan 2.2 scheduler-transition curvature discontinuity) are validity/generalization concerns rather than circularity. Absence of a step-size error analysis weakens the theoretical support but does not reduce any result to its inputs.
Axiom & Free-Parameter Ledger
free parameters (7)
- learning stabilizer EMA beta =
0.9985 (FLUX.1-dev), 0.995 (Qwen-Image, Wan 2.2)
- curvature_scale =
2.0 (default)
- adaptive gate tolerance =
not specified
- anchor_interval =
4
- max_consecutive_skips =
2
- protected head/tail windows =
0-1 head, 1 tail depending on workflow
- validation thresholds =
absolute floor 1e-8, relative floor 1e-6, RES cap 50x
axioms (4)
- domain assumption Epsilon is locally polynomial to order 2-4 in step index over the skip window.
- domain assumption SSIM >= 0.95 corresponds to perceptual equivalence or 'imperceptible' differences.
- domain assumption Single-seed, single-run comparisons are representative of quality and timing.
- domain assumption Replacing the model output with an extrapolated denoised value preserves the sampler's numerical update validity.
Cite this review
Pith. "Pith review of FSampler: Training Free Acceleration of Diffusion Sampling via Epsilon Extrapolation." pith.science (2026). https://pith.science/paper/GZP4JU7U
@misc{pith2026251109180,
author = {Pith},
title = {Pith review of: FSampler: Training Free Acceleration of Diffusion Sampling via Epsilon Extrapolation},
year = {2026},
howpublished = {\url{https://pith.science/paper/GZP4JU7U}},
note = {Machine review of arXiv:2511.09180}
}
read the original abstract
FSampler is a training free, sampler agnostic execution layer that accelerates diffusion sampling by reducing the number of function evaluations (NFE). FSampler maintains a short history of denoising signals (epsilon) from recent real model calls and extrapolates the next epsilon using finite difference predictors at second order, third order, or fourth order, falling back to lower order when history is insufficient. On selected steps the predicted epsilon substitutes the model call while keeping each sampler's update rule unchanged. Predicted epsilons are validated for finiteness and magnitude; a learning stabilizer rescales predictions on skipped steps to correct drift, and an optional gradient estimation stabilizer compensates local curvature. Protected windows, periodic anchors, and a cap on consecutive skips bound deviation over the trajectory. Operating at the sampler level, FSampler integrates with Euler/DDIM, DPM++ 2M/2S, LMS/AB2, and RES family exponential multistep methods and drops into standard workflows. FLUX.1 dev, Qwen Image, and Wan 2.2, FSampler reduces time by 8 to 22% and model calls by 15 to 25% at high fidelity (Structural Similarity Index (SSIM) 0.95 to 0.99), without altering sampler formulas. With an aggressive adaptive gate, reductions can reach 45 to 50% fewer model calls at lower fidelity (SSIM 0.73 to 0.74).
Figures
Reference graph
Works this paper leans on
-
[1]
Lu, C., Zhou, Y., Bao, F., Chen, J., Li, C. and Zhu, J. (2022a) 'DPM-Solver: A Fast ODE Solver for Diffusion Probabilistic Model Sampling in Around 10 Steps', Advances in Neural Information Processing Systems (NeurIPS 2022). Available at: https://arxiv.org/abs/2206.00927
Pith/arXiv arXiv 2022
-
[2]
Lu, C., Zhou, Y., Bao, F., Chen, J., Li, C. and Zhu, J. (2022b) 'DPM-Solver++: Fast Solver for Guided Sampling of Diffusion Probabilistic Models'. arXiv:2211.01095. Available at: https://arxiv.org/abs/2211.01095
-
[3]
Wang, G., Cai, Y., Li, L., Peng, W. and Su, S. (2025) 'PFDiff: Training-Free Acceleration of Diffusion Models Combining Past and Future Scores', International Conference on Learning Representations (ICLR 2025). Available at: https://arxiv.org/abs/2408.08822
Pith/arXiv arXiv 2025
-
[4]
Liu, F., Zhang, S., Wang, X., Wei, Y., Qiu, H., Zhao, Y., Zhang, Y., Ye, Q. and Wan, F. (2025) 'Timestep Embedding Tells: It's Time to Cache for Video Diffusion Model', IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR 2025). Available at: https://arxiv.org/abs/2411.19108
Pith/arXiv arXiv 2025
-
[5]
Zhao, W., Bai, L., Rao, Y., Zhou, J. and Lu, J. (2023) 'UniPC: A Unified Predictor-Corrector Framework for Fast Sampling of Diffusion Models', Advances in Neural Information Processing Systems (NeurIPS 2023). Available at: https://arxiv.org/abs/2302.04867
Pith/arXiv arXiv 2023
-
[6]
Zhang, Q. and Chen, Y. (2022) 'Fast Sampling of Diffusion Models with Exponential Integrator', International Conference on Learning Representations (ICLR 2023). Available at: https://arxiv.org/abs/2204.13902
Pith/arXiv arXiv 2022
-
[7]
Song, J., Meng, C. and Ermon, S. (2021) 'Denoising Diffusion Implicit Models', International Conference on Learning Representations (ICLR 2021). Available at: https://arxiv.org/abs/2010.02502
Pith/arXiv arXiv 2021
-
[8]
and Laine, S
Karras, T., Aittala, M., Aila, T. and Laine, S. (2022) 'Elucidating the Design Space of Diffusion-Based Generative
2022
-
[9]
Available at: https://arxiv.org/abs/2206.00364
Models', Advances in Neural Information Processing Systems (NeurIPS 2022). Available at: https://arxiv.org/abs/2206.00364
Pith/arXiv arXiv 2022
-
[10]
Zhang, Q., Song, J. and Chen, Y. (2023) 'Improved Order Analysis and Design of Exponential Integrator for Diffusion Models Sampling'. arXiv:2308.02157. Available at: https://arxiv.org/abs/2308.02157 Models Used in Experiments Black Forest Labs (2024) FLUX.1 [dev]. Available at: https://huggingface.co/black-forest-labs/FLUX.1-dev (Accessed: October 2025). ...
Pith/arXiv arXiv 2023
This paper was first reviewed by deepseek-v4-flash on August 3, 2026.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.