acceptodds
Under review as a conference paper at ICLR 2027

Endpoint-Complete Diffusion Sampling via Log-SNR Boundary Closures

Abstract

Modern fast diffusion samplers are typically designed as finite-interval integrators over a prescribed sequence of noise levels. This view reduces discretization error between neighboring nodes, but leaves a separate source of error at the two truncated ends of the sampling trajectory. We revisit diffusion sampling in log-SNR space, where the complete trajectory extends from the low-SNR limit to the high-SNR limit . Taking these limits reveals an asymmetric boundary structure: the low-SNR endpoint admits a regular signal-prediction closure, whereas the high-SNR endpoint admits a regular noise-prediction closure. Motivated by this asymmetry, we propose endpoint-complete diffusion sampling, which keeps the interior solver unchanged and replaces only the first and last updates with endpoint-specific boundary plug-ins: DPM-FS and DPM-LS. We further connect the same endpoint asymmetry to generalized velocity prediction and show that, under variance normalization, EDM-style residual preconditioning maps non-degenerate linear targets to the same generalized velocity direction up to sign. Experiments on CIFAR-10, AFHQv2, FFHQ, and ImageNet demonstrate improved few-step sampling quality.

Primary area:
generative models
Keywords:
Diffusion Sampling, Endpoint Closures, Time-Step Scheduling

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_3QEGOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes second-order updates for the beginning and end of diffusion sampling, together with schedules that allocate more of the trajectory to these updates. It combines them with existing solvers and uses a mean estimate, or a model prediction in the SDXL experiments, to initialize sampling. Results across several models show improved sample quality at low NFE.

Critical strengths
  1. The method works with existing solvers and pretrained models. When a mean estimate is available, the second-order first step requires only one evaluation of the main model. This is useful when the sampling budget is small.

  2. The experiments show gains across multiple models and solvers. Improvements with the analytic Extend schedule (Tables 3–6) are particularly reassuring, since they do not rely on the modified GITS search.

Critical weaknesses
  1. The methodological novelty is less clear to me. LS-1 is the usual final denoising step, while FS-2 and LS-2 have a two-stage Runge–Kutta form after the variable changes in Appendix A. The paper could explain more clearly what the endpoint analysis and initialization add to these existing ideas.

  2. The gain from the boundary updates is difficult to separate from that of the schedule. Figure 14 compares FS-only and FSLS at equal NFE, but gives them different freedom to optimize the final interval. The additional LS gains are also mixed, leaving uncertainty about the benefit of the last-step update itself.

Questions for authors
  1. Could you explain the relationship between FS-2/LS-2 and existing second-order integration schemes, and identify the contribution of the endpoint formulation and initialization?

  2. At equal NFE, could you isolate the gains from LS-2 and from FS-2 without a precomputed mean, using baselines with comparable timestep optimization?

3 comments

We thank the reviewer for the careful assessment and the specific questions about the relationship between our endpoint formulas and existing second-order methods, as well as the individual contributions of FS2 and LS2. We further clarify the roles of endpoint analysis and initialization and provide ablations without a precomputed mean at matched NFE and with comparable timestep optimization.

W1 & Q1. Relationship to Second-Order Integration and the Contribution of Endpoint Initialization

Could you explain the relationship between FS-2/LS-2 and existing second-order integration schemes, and identify the contribution of the endpoint formulation and initialization?

FS2 and LS2 have a two-stage RK2 structure under appropriate changes of variables. Our endpoint analysis identifies the coordinates and prediction forms for this structure:

=D_\theta(z_\lambda,\lambda),\qquad \frac{d(z_\lambda/\alpha_\lambda)}{d(e^{-\lambda})} =\epsilon_\theta(z_\lambda,\lambda).$$ The former provides a signal-form closure at the low-SNR endpoint, while the latter provides a noise-form closure at the high-SNR endpoint. **This asymmetry determines which quantities to integrate, which boundary conditions to use, and how to avoid singular normalization at the two endpoints.** Proposition 1 further establishes the corresponding higher-order endpoint expansions and error orders. LS1 corresponds to conventional terminal denoising as the first-order case of this framework. For the first step, the analysis identifies the required signal anchor, whose ideal zero-SNR limit is the conditional mean. EFS provides this initialization without an additional main-network evaluation; alternatively, it can be replaced with a network prediction whose evaluation is counted toward the NFE budget. The ablations below use the latter approach to directly examine the benefits of FS2/LS2 without a precomputed mean.

W2. The Contribution of LS2 under a Fixed Schedule

The additional LS gains are also mixed, leaving uncertainty about the benefit of the last-step update itself.

We have added ablations using the same timestep nodes without a precomputed mean and find that adding the LS module can affect sample quality in certain settings. In models such as RAE and PixelDiT, this issue is caused by the “curse of dimensionality” encountered when directly predicting noise in very high-dimensional spaces at high SNR [1]. To investigate this issue, we additionally conducted experiments with JiT-L/32 and PixelDiT2 [2]. Both models use the v-loss discussed in our paper, but their networks predict the signal. Consequently, they can accurately estimate high-dimensional noise even at high SNR, allowing the LS module to improve sample quality on these models. We will include a more detailed discussion of this behavior in the paper.

The four values in each cell correspond to DDIM / DPM-Solver++2M / UniPC / iPNDM-V, respectively. FSLS uses FS cache where applicable, and the total budget is 5 NFE. For the initial signal estimate, we use one network evaluation to obtain the signal anchor, following an approach similar to Table 2 in the paper by choosing a sufficiently small λ\lambda. This evaluation is included in the NFE budget.

FS + LSFSLS
CIFAR-109.84/10.80/7.67/9.7420.16/14.29/9.85/14.6812.94/15.44/10.98/12.84
AFHQv26.49/5.68/5.19/6.229.88/8.34/8.43/8.1411.72/10.05/9.73/10.89
FFHQ12.12/11.05/15.49/12.6018.45/12.03/15.93/14.6422.07/20.56/23.03/23.41
EDM2-S10.46/8.52/8.15/9.6618.56/9.81/9.30/10.9118.14/13.59/11.79/16.41
RAE12.02/13.68/12.05/12.069.53/12.47/11.38/9.3617.79/20.25/19.72/17.29
PixelDiT13.43/11.70/12.51/12.4312.12/9.99/9.98/10.0816.77/15.93/16.19/15.60
Additional Models
JiT-L/3211.29/10.29/10.00/10.7115.67/12.45/12.16/12.8812.51/11.66/11.37/12.31
PixelDiT2-H/1610.16/9.20/8.63/10.8613.30/11.54/10.65/12.4311.03/10.11/9.37/11.40

Q2. Ablations with Independently Optimized Schedules

At equal NFE, could you isolate the gains from LS-2 and from FS-2 without a precomputed mean, using baselines with comparable timestep optimization?

We have added ablations with independently optimized timestep nodes and without a precomputed mean.

FS + LS w/ EFSFS + LSFSLS
CIFAR-1010.53/8.37/4.68/8.659.84/10.80/7.67/9.7415.73/9.66/10.06/11.6511.39/10.37/9.90/11.06
AFHQv27.86/6.61/5.27/5.376.49/5.68/5.19/6.229.11/7.45/7.39/7.4510.06/7.96/8.69/9.10
FFHQ13.58/11.26/8.91/11.9812.12/11.05/15.49/12.6017.39/20.78/17.36/15.3419.56/17.10/16.72/18.62
EDM2-S7.94/5.65/7.31/5.8710.46/8.52/8.15/9.6618.16/8.11/7.08/13.0514.24/11.27/10.40/15.45
RAE10.02/10.03/11.05/10.9912.02/13.68/12.05/12.0611.54/18.42/11.89/11.7419.94/17.02/16.88/15.09
PixelDiT9.39/7.95/7.36/8.7713.43/11.70/12.51/12.4311.51/10.21/11.05/10.5317.80/16.23/15.71/16.13
Additional Models
JiT-L/329.56/7.50/7.00/8.2511.29/10.29/10.00/10.7114.47/11.32/10.93/11.6112.62/10.63/10.99/11.44
PixelDiT2-H/167.23/6.46/6.01/7.3110.16/9.20/8.63/10.8613.61/11.58/10.38/11.1611.10/9.53/9.87/10.22

Without a precomputed mean, at matched NFE and with comparable timestep optimization, FS+LS outperforms LS-only in 30/32 configurations, demonstrating the additional benefit of FS2.

Across all 8 configurations of JiT-L/32 and PixelDiT2, FS+LS outperforms FS-only. This holds under both fixed and independently optimized schedules, further demonstrating the benefit of LS2.

References

[1] Li, T., et al. Back to Basics: Let Denoising Generative Models Denoise. CVPR 2026.

[2] Yu, Y., et al. PixelDiT2: Representation-Grounded Pixel Diffusion Transformers. arXiv 2026.

Thank you for the additional explanations and comparisons at the same sampling budget. They address my main empirical concern by showing that FS-2 generally improves FID even without a precomputed mean, while the gains from LS-2 vary across models. My overall assessment remains positive, and I will keep my score at 3.

Reviewer_pj5ZOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies boundary errors in diffusion sampling and proposes asymmetric treatments for the low- and high-SNR endpoints. DPM-FS constructs the initial finite state using signal prediction and an estimated signal anchor, typically a dataset or conditional mean, while DPM-LS completes sampling using noise prediction. These modules are combined with existing interior solvers and schedules that allocate evaluations between the boundaries and interior. The paper also connects endpoint-compatible prediction to generalized velocity and residual preconditioning. Experiments across several solvers, datasets, and model families demonstrate improvements in few-step sampling.

Critical strengths
  1. The paper provides a coherent formulation of boundary treatment as part of diffusion sampler design. The complementary roles of signal prediction at low SNR and noise prediction at high SNR are clearly motivated, and the boundary expansions appear sound under the stated assumptions. The distinction between the network's residual target, its exposed prediction, and the quantity used by the solver is also useful.

  2. The main experimental comparisons cover multiple solvers and model families at matched total NFE. Figure 3 and Tables 3–6 explicitly compare Polynomial and Extend with and without FSLS, providing evidence that endpoint treatment contributes beyond changing the scheduling rule alone. The inclusion of uncached variants and configurations where FSLS degrades performance makes the interaction between boundary treatment and node allocation more informative.

Critical weaknesses
  1. The GITS-based comparisons do not fully isolate the contribution of FSLS under comparable schedule calibration. Although the Polynomial/Extend ablations provide useful evidence, Appendices D.3 and E.2 indicate that ordinary GITS and endpoint-enabled GITS use different optimization objectives and search procedures. The latter calibrates the complete sampler using terminal-image feature matching, with a higher offline calibration cost in the reported setting. Readjusting schedules is reasonable when boundary updates change evaluation allocation, but the GITS comparison alone does not establish how much of the gain remains against a sampler without FSLS calibrated using the same objective and comparable search budget. Such a control would strengthen the attribution of the GITS-based gains.
Questions for authors
  1. Table 1 and Tables 7–8 already include the flow-matching backbones RAE and PixelDiT. Could the authors clarify which endpoint errors identified in the paper persist under direct native-time velocity integration, and whether the reported baselines already cover or are equivalent to this setting? For example, DDIM can coincide with native-time Euler for a linear flow path at the same time nodes. If standard native-time sampling configurations are not already represented, a comparison using the same checkpoints, matched total NFE, and comparably calibrated schedules would clarify the additional benefit of the proposed boundary closures.

  2. Could the authors calibrate a sampler without FSLS using the same terminal-feature objective, teacher outputs, calibration data, and a comparable offline search budget as endpoint-enabled GITS? How much of the GITS-based improvement remains under this comparison?

1 comment

We thank the reviewer for the careful assessment and the specific questions about schedule calibration and native-time velocity integration. We have added controlled comparisons under matched calibration conditions and further clarify the relationship between native-time velocity integration and our endpoint formulas.

W1 & Q2. Comparisons under Matched Calibration

Could the authors calibrate a sampler without FSLS using the same terminal-feature objective, teacher outputs, calibration data, and a comparable offline search budget as endpoint-enabled GITS? How much of the GITS-based improvement remains under this comparison?

We have added this control. For clarity, we denote the optimized version of GITS by GITS*.

The four values in each cell correspond to DDIM / DPM-Solver++2M / UniPC / iPNDM-V, respectively. FSLS uses FS cache where applicable, and all configurations use a total budget of 5 NFE.

GITSGITS*GITS* + FSLS
CIFAR-1030.07/14.58/13.95/10.2317.16/9.82/8.97/11.8810.53/8.37/4.68/8.65
AFHQv219.70/13.50/12.26/9.8512.30/8.50/8.18/9.017.86/6.61/5.27/5.37
FFHQ42.01/17.74/16.74/15.9326.59/14.98/13.59/15.9113.58/11.26/8.91/11.98
EDM2-S54.83/33.63/30.11/13.0324.01/9.18/7.54/10.447.94/5.65/7.31/5.87
RAE49.36/53.17/52.24/51.8730.76/26.94/26.44/30.3610.02/10.03/11.05/10.99
PixelDiT19.58/20.73/20.87/19.8812.63/10.72/11.31/13.679.39/7.95/7.36/8.77

FSLS outperforms GITS* under matched calibration in all 24 model–solver combinations. This shows that endpoint treatment retains clear benefits even when the ordinary sampler receives comparable timestep optimization.

Q1. Relation to Native-Time Velocity Integration

Which endpoint errors identified in the paper persist under direct native-time velocity integration?

For a linear flow with a bounded native velocity predictor, direct integration avoids the singular error amplification associated with incompatible signal/noise conversions. The first and last finite intervals still require numerical approximation. In particular, native-time Euler provides precisely the first-order endpoint treatment described by our FS1 and LS1.

Consider the noise-to-data parameterization zt=(1−t)ϵ+txz_t=(1-t)\epsilon+tx. The signal prediction associated with the velocity field is Dθ(zt,t)=zt+(1−t)vθ(zt,t).D_\theta(z_t,t)=z_t+(1-t)v_\theta(z_t,t). The first and last Euler steps therefore satisfy

z_1=z_{t_R}+(1-t_R)v_\theta(z_{t_R},t_R)=D_\theta(z_{t_R},t_R).$$ These are precisely FS1 with a network-predicted signal anchor and LS1 with conventional terminal denoising. Thus, native-time Euler already uses endpoint-compatible first-order updates. **FS2 and LS2 extend these boundary treatments to second order while retaining the existing interior solver.** This addresses the discretization of the boundary intervals even when the native velocity parameterization itself is regular. For the same velocity field and corresponding finite time nodes, the DDIM update also satisfies $$z_{t'}=z_t+(t'-t)v_\theta(z_t,t).$$ Thus, our existing DDIM baselines already cover native-time Euler updates over the corresponding finite intervals.
Reviewer_hm2ROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

In this paper the authors proposed a method for completing boundary points of diffusion trajectories for fast samplers. The study appears to be technically sound. However, the proposed method increases model complexity. Whether the added complexity is justified in terms of performance gain remains unclear. Second, the compatibility of the proposed method with existing samplers remains unclear. Overall, I am not yet convinced that the proposed method represents a substantial contribution to the field of diffusion models.

Critical strengths
  1. The writing of the paper can generally be followed without too much difficulty, although there are places that require clarification.

  2. The experimental validation of the proposed method seems extensive.

Critical weaknesses
  1. On page 2, line 070, the authors wrote "Conventional sampling schedules are designed for finite-interval interior solvers, with the sampling endpoints typically fixed by the schedule itself and no explicit allocation to endpoint solvers."

The proposed boundary completion method adds complexity. Table 1 and 2 show performance gains, but the FIDs overall are much worse than SOTA. Is the method designed specifically for fast samplers under low NFE? As shown in the supplementary tables 3,4,5,6, when the number of NFE increases, the performance gain by the proposed method decreases. It is also unclear how compatible is the proposed method with existing samplers. Can the proposed method be easily integrated?

Overall, it remains unclear whether the research contribution of this study is sufficiently substantial. Other than Figure 1, please provide a real-world example demonstrating that without the proposed boundary completion, how the quality of generative output will deteriorate.

  1. Page 1, line 72, what does "FSLS" stand for? Please spell it out the first time this abbreviation is used.
Questions for authors
  1. What are the strategies used by existing fast diffusion samplers for boundary completion? This is not entirely clear from the paper.

  2. How compatible is the proposed method with existing fast diffusion samplers?

  3. The FID scores in Table 1 and Table 2 appear to be substantially worse than SOTA (e.g. for EDM). Please clarify.

  4. What are the limitations of this study?

2 comments

We thank the reviewer for the assessment and the opportunity to clarify our contributions, evaluation budgets, compatibility, and limitations.

W1 & Q3. FID Comparison with SOTA

“The FIDs overall are much worse than SOTA. Is the method designed specifically for fast samplers under low NFE?”

We clarify that the results in Tables 1 and 2 were obtained with budgets of 5 and 10 NFE (Number of Function Evaluations), respectively. Our method primarily targets fast sampling under low-NFE budgets.

For example, the FID of 2.56 reported for EDM2-S in the original work was obtained using 63 NFE by using Heun2nd solver and polynomial schedule. At the same total budget of 5 NFE, the Heun2nd + Polynomial baseline in Table 1 yields an FID of 208.43, whereas Extend + FSLS and GITS + FSLS achieve FIDs of 7.41 and 8.54, respectively.

As the total NFE increases, the proportion of NFE accounted for by our endpoint sampler decreases, and at this point, the improvement in sampling quality provided by the internal solver gradually becomes the dominant factor.

W1. Sampling Defects Without Endpoint Treatment

“Other than Figure 1, please provide a real-world example demonstrating that without the proposed boundary completion, how the quality of generative output will deteriorate.”

Inaccurate treatment of the low-SNR boundary can distort low-frequency image statistics, including global brightness and channel means. Such effects have been documented in prior work [1–2]. Figure 5 illustrates this issue, especially in the example with the prompt “A bald eagle against a white background.” The original SDXL configuration without endpoint solvers produces a visibly gray background instead of the requested white background, along with lower overall brightness.

Mismatched prediction forms at low SNR can amplify signal-recovery errors and cause over- or underexposure, as illustrated in Figure 16(a–b).

High-SNR endpoint errors can affect fine image details that are difficult to discern visually. Figure 14 shows that FSLS generally improves FID over FS alone at matched NFE, particularly in few-step sampling.

W2. Meaning of FSLS

“What does ‘FSLS’ stand for?”

Thank you for pointing out this omission. Here, “with FSLS” means “solvers with DPM-FS and DPM-LS,” namely, standard interior solvers combined with our proposed solvers for the two endpoints. We will spell out this abbreviation at its first occurrence.

Q1. Related Work

“What are the strategies used by existing fast diffusion samplers for boundary completion?”

To the best of our knowledge, there are currently no fast diffusion solvers for boundary completion in the literature. Related approaches such as AFS [3] and OMS [4] address initialization through analytical or learned first-step treatments. Our method derives dedicated higher-order formulas for both endpoints and combines them with existing interior solvers.

Q2. Compatibility with Fast Interior Solvers

“How compatible is the proposed method with existing fast diffusion samplers? Can the proposed method be easily integrated?”

Table 1 and Tables 3–8 already demonstrate combinations of our endpoint solvers with six interior solvers. At finite transition nodes, the states and predictions involved in the endpoint updates can be converted, according to the corresponding parameterization, into the signal, noise, or velocity representations required by the interior solver.

Furthermore, we use half-log-SNR as a common coordinate. The endpoint formulas are determined by the signal and noise mixing coefficients and do not depend on a particular time parameterization. Evaluating a pretrained model only requires mapping this coordinate to the model's native time or noise level. For multistep solvers that support history caching, the combined first-step prediction can also be used to initialize the cache.

Q4. Limitations

“What are the limitations of this study?”

Our method requires the model to reliably support different prediction forms at the two endpoints, namely, signal prediction in the low-SNR region and noise prediction in the high-SNR region. If the model cannot provide accurate and numerically stable predictions in the corresponding region, the accuracy of the endpoint solver is also limited, and the solver may fail in severe cases. Our analysis of generalized velocity and residual preconditioning serves to clarify this applicability condition.

References

[1] Diffusion with Offset Noise.

[2] Lin, S., et al. Common Diffusion Noise Schedules and Sample Steps Are Flawed. WACV 2024.

[3] Dockhorn, T., et al. GENIE: Higher-Order Denoising Diffusion Solvers. NeurIPS 2022.

[4] Hu, M., et al. One More Step: A Versatile Plug-and-Play Module for Rectifying Diffusion Schedule Flaws and Enhancing Low-Frequency Controls. CVPR 2024.

We would appreciate any further feedback, particularly on whether our responses have addressed your main concerns. Thank you for your time.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.