acceptodds
Under review as a conference paper at ICLR 2027

Mean-Shifted Intermediate Sampling: Training-Free Acceleration for Few-Step Diffusion Generation

Abstract

Few-step diffusion sampling suffers from severe discretization errors due to large step sizes. Some existing methods reduce the average step size by starting reverse sampling from an intermediate distribution, but they often require training additional models or rely on real samples during inference. In this paper, we propose Mean-Shifted Intermediate Sampling (MSIS), a simple training-free strategy that approximates the high-noise intermediate distribution with mean-shifted Gaussian noise. The mean shift is determined by the forward noising formula and the dataset itself, so it can be computed once offline and reused across samplers, timestep schedules, and models sharing the same input space, with negligible inference overhead. We further show that near the Gaussian endpoint, the mean shift is the dominant data-dependent deviation, while covariance and higher-order effects are weaker, justifying mean-shifted Gaussian noise as a first-order approximation. Extensive experiments demonstrate the effectiveness and generality of MSIS across datasets, samplers, schedules, and architectures. On conditional ImageNet \(256\times256\), MSIS requires only \(2.22\) minutes of one-time preprocessing on a single RTX 3090 and reduces FID by \(35.18%\)–\(55.33%\) across the evaluated DDIM step budgets. Our code will be made publicly available upon acceptance.

Primary area:
generative models
Keywords:
Diffusion Models, Accelerated Sampling, Truncated Reverse Diffusion

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_tbbhOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

The authors propose MSIS, which improves the initial noise to enable diffusion models to generate better images with very few denoising steps. The core approach is to precompute the mean of 1% of the training images, select an intermediate noise starting point, add the scaled mean to newly sampled random noise, and use the original model to complete the remaining denoising steps, thereby accelerating training. The approach is validated on the ImageNet dataset using the SIT model.

Critical strengths
  1. The paper is clearly written and easy to follow.
  2. The experiments are reasonably thorough and cover several datasets and samplers.
Critical weaknesses
  1. The text-to-image evaluation is narrow: Stable Diffusion 2.1 on MS-COCO alone is not enough. Does the method still help FLUX.2, Z-Image, or Qwen-Image on complex prompts?
  2. The method may sacrifice diversity by discarding data covariance. For a VP path, the true intermediate covariance is (a_t^2I+b_t^2\Sigma), whereas MSIS uses (I). Since at2+bt2=1a_t^2+b_t^2=1, MSIS underestimates variance by bt2(λ−1)b_t^2(\lambda-1) along any data-covariance eigenvector with eigenvalue λ>1\lambda>1. Under a locally linear denoising map F(x)≈Jx+dF(x)\approx Jx+d, define output diversity as the expected squared distance between independent outputs. Then D_{\mathrm{true}}-D_{\mathrm{MSIS}}\ \=2b_t^2\operatorname{tr}!\left[J(\Sigma-I)J^\top\right].\ This is positive when the retained directions are dominated by data variance above one. Thus, diversity loss is a concrete risk, although not an unconditional result for nonlinear denoising. The full-covariance variant also improves recall from 0.503 to 0.536. Why do the authors not directly test diversity across seeds for the same prompt?
  3. The choice of z over a_{t_s}z is not tested. The authors argue that scaling the noise causes oversmoothing, but provide no direct comparison. This is a central design choice, not a detail that can be justified by intuition alone.
  4. Lower FID does not establish better handling of complex prompts, and average CLIP Score offers limited evidence. Does MSIS improve composition, text rendering, and instruction following on Qwen-Image-Bench (Text-to-Image)?
  5. The experiments leave out models already trained for few-step generation. Does MSIS offer any additional benefit on DMD/LCM-distilled models? Its applicability to MeanFlow and GAN-based generators is also unclear. Without these results, the practical value of the acceleration claim is difficult to judge.
Questions for authors

No

10 comments
AuthorsOpenReview ↗

Response to Q1 & Q4-1: Generalization to Z-Image and Fine-Grained Text-to-Image Evaluation

Thank you for raising these important concerns regarding the generalization of MSIS to other large-scale text-to-image models and its impact on generation quality under complex prompts.

First, we would like to clarify the primary objective of MSIS. Our method is not designed to directly enhance fine-grained generation capabilities such as composition, text rendering, or instruction following. Instead, its main objective is to reduce the discretization errors introduced by excessively large sampling steps in few-step generation. Since these fine-grained capabilities can also be affected by discretization errors, reducing such errors may indirectly alleviate the degradation of these capabilities in few-step generation, or at least avoid introducing additional degradation due to the initialization strategy.

To further investigate this effect and address the reviewer's concerns about cross-model generalization, we conduct additional experiments using ​Z-Image on Qwen-Image-Bench. This allows us to evaluate MSIS on another large-scale text-to-image model and examine its effects on fine-grained generation capabilities.

Experimental Setup

Qwen-Image-Bench contains 1,000 prompts and provides a hierarchical evaluation framework covering various aspects of text-to-image generation. However, both Z-Image generation and benchmark evaluation involve substantial computational costs. Given the limited computational budget and hardware resources available during the rebuttal period, we adopt a reduced evaluation protocol.

Specifically, we randomly select 500 prompts from the official 1,000-prompt benchmark and focus on three L1 dimensions most relevant to the reviewer's concerns:

  • Aesthetics: Evaluates overall visual aesthetic quality, including composition, color harmony, lighting, anatomical portraiture, emotional expression, and style control. In particular, the Composition subcategory directly evaluates the visual composition of generated images.
  • Creative Generation: Evaluates creative generation capabilities, including imagination, feature matching, logical resolution, text rendering, design applications, and visual storytelling. The Text Rendering subcategory specifically evaluates text accuracy, text layout, font rendering, and cross-lingual text generation.
  • Alignment: Evaluates prompt-image consistency through attributes, actions, spatial layout, object relations, and scene-level requirements. This dimension serves as the closest benchmark-level proxy for instruction following.

Regarding MSIS initialization, since the original training dataset of Z-Image is not available to us for computing the global data mean, we randomly select an additional set of ​100 prompts, which are completely disjoint from the 500 evaluation prompts. We reverse-sample these 100 prompts to the 0.98 point and use the resulting samples to estimate the mean for MSIS initialization.

We compare the original Z-Image initialization (​Raw​) with our MSIS initialization (​Ours​). ​Both configurations use exactly 10 sampling steps, ensuring that the sampling budget remains identical and that the only difference is the initialization strategy. All other generation settings follow the corresponding official Z-Image configurations and are kept consistent between the two configurations.

For evaluation, we follow the Qwen-Image-Bench evaluation protocol, aggregating scores hierarchically from L3 evaluation facets to L2 subcategories and then to L1 dimensions. N/A entries are excluded from score aggregation.

Overall Results

The overall results for the three selected L1 dimensions are summarized below.

Table 1: Overall comparison between Raw and Ours on Qwen-Image-Bench.

L1 DimensionRawOursPaired Δ
Aesthetics46.2946.97+0.67
Creative Generation39.0639.87+0.80
Alignment49.7549.94+0.19

The results show positive changes across all three evaluated L1 dimensions. Specifically, Aesthetics, Creative Generation, and Alignment exhibit paired improvements of 0.67, 0.80, and 0.19 points, respectively.

These results provide further evidence of the cross-model generalization of MSIS to Z-Image. More importantly, under the same 10-step sampling budget, introducing MSIS does not degrade the overall scores of these three dimensions. Instead, all three exhibit positive changes, suggesting that MSIS can preserve overall generation quality in complex text-to-image tasks.

To further investigate the effects on the fine-grained capabilities raised by the reviewer, we examine the corresponding L2 subcategories.

AuthorsOpenReview ↗

Response to Q1 & Q4-2: Generalization to Z-Image and Fine-Grained Text-to-Image Evaluation

Detailed Results and Analysis

(a) Aesthetics: Visual Composition and Aesthetic Quality

The detailed results for Aesthetics are presented below.

Table 2: Detailed comparison of Aesthetics subcategories.

SubcategoryRawOursPaired Δ
Aesthetics (L1)46.2946.97+0.67
Composition46.0644.55-1.52
Color Harmony60.6160.71+0.10
Lighting52.9953.81+0.81
Anatomical Portraiture6.595.29-0.99
Emotional Expression48.3149.39+0.88
Style Control40.6244.90+4.17

(b) Creative Generation: Text Rendering and Creative Generation Capabilities

The detailed results for Creative Generation are presented below.

Table 3: Detailed comparison of Creative Generation subcategories.

SubcategoryRawOursPaired Δ
Creative Generation (L1)39.0639.87+0.80
Imagination57.6960.00+2.61
Feature Matching43.9344.55+0.43
Logical Resolution37.4134.04-1.03
Text Rendering29.7030.27+0.57
Design Applications42.1641.98-0.35
Visual Storytelling55.8655.54-0.25

(c) Alignment: Instruction Following and Prompt-Image Consistency

The detailed results for Alignment are presented below.

Table 4: Detailed comparison of Alignment subcategories.

SubcategoryRawOursPaired Δ
Alignment (L1)49.7549.94+0.19
Attributes56.5956.20-0.39
Actions36.1036.11+0.98
Layout50.4252.04+1.47
Relations40.4441.90+1.46
Scene60.7859.78-1.01

Across the three dimensions, we observe positive changes in several fine-grained capabilities relevant to the reviewer's concerns. In particular, Text Rendering exhibits a positive change of +0.57, while Layout and Relations show improvements of +1.47 and +1.46, respectively. Meanwhile, Composition exhibits a decrease of -1.52, and some other subcategories also show negative changes. Nevertheless, all three L1 dimensions exhibit positive paired differences, indicating that MSIS does not lead to an overall degradation in these evaluated capabilities.

Discussion

We emphasize that MSIS is not designed to directly enhance composition, text rendering, or instruction following. Its primary objective is to reduce discretization errors introduced by large sampling steps in few-step generation. Since these fine-grained capabilities can also be affected by discretization errors, MSIS may indirectly alleviate their degradation by reducing such errors.

Our experimental results are consistent with this expectation. Under the same 10-step sampling budget, all three evaluated L1 dimensions exhibit positive paired differences. Furthermore, Text Rendering, Layout, and Relations show positive changes, suggesting that MSIS can help preserve these fine-grained capabilities and mitigate potential degradation in certain aspects. Although some individual subcategories still exhibit decreases, the overall results do not indicate systematic degradation caused by MSIS.

We acknowledge that, due to the limited computational budget during the rebuttal period, our additional evaluation covers only 500 of the 1,000 official prompts and focuses on three L1 dimensions most relevant to the reviewer's concerns. Further evaluations on models such as FLUX.2 and Qwen-Image, as well as a comprehensive evaluation across all benchmark dimensions, would be valuable future work. We will incorporate these additional results and clarify the corresponding experimental settings and limitations in the revised manuscript.

We thank the reviewer again for these valuable suggestions, which have helped us further validate the applicability of MSIS across different text-to-image models and better understand its effects on complex text-to-image generation.

Note: Paired Δ is computed by averaging per-prompt score differences over prompts with valid scores under both configurations. Since the Raw and Ours scores may be aggregated over different sets of valid prompts, Paired Δ may differ from the difference between the two reported average scores.

AuthorsOpenReview ↗

Response to Q2-1: Potential Impact of Covariance Approximation on Generation Diversity

Thank you for raising the concern that approximating the intermediate covariance with an isotropic Gaussian may sacrifice generation diversity. We address this concern from both theoretical and empirical perspectives.

Theoretical Analysis: Covariance Mismatch Does Not Necessarily Imply Reduced Output Diversity

We first clarify that an initialization-level covariance mismatch does not necessarily translate into reduced diversity in the final generated images.

Consider the VP diffusion process:

\qquad \epsilon\sim\mathcal N(0,I),$$ where $x_0$ and $x_1$ are independent, and the data distribution has covariance $\Sigma=\operatorname{Cov}(x_1)$. The covariance of the true intermediate distribution is $$\Sigma_t^{\mathrm{true}} =a_t^2I+b_t^2\Sigma.$$ In contrast, MSIS uses an isotropic initialization with covariance $$\Sigma_t^{\mathrm{MSIS}}=I.$$ Using the VP condition $a_t^2+b_t^2=1$, their covariance difference is $$\Delta\Sigma_t =b_t^2(\Sigma-I).$$ Therefore, MSIS indeed underestimates the initialization variance along data-covariance eigenvectors with eigenvalues greater than one, while overestimating it along directions with eigenvalues smaller than one. However, this covariance mismatch alone does not establish that the final generated images necessarily exhibit reduced diversity. To understand this distinction, consider a local linear approximation of the denoising mapping: $$F(x)\approx Jx+d,$$ where $J$ denotes the local Jacobian of the denoising mapping. Defining output diversity as the expected squared distance between two independent outputs, we have $$D=\mathbb E\left[ \|F(x^{(1)})-F(x^{(2)})\|_2^2 \right] =2\operatorname{tr}(J\Sigma_xJ^\top).$$ Consequently, the difference in output diversity between the true intermediate distribution and MSIS is approximately $$\begin{aligned} D_{\mathrm{true}}-D_{\mathrm{MSIS}} &=2\operatorname{tr} \left[J(\Sigma_t^{\mathrm{true}}-I)J^\top\right]\\ &=2b_t^2\operatorname{tr} \left[J(\Sigma-I)J^\top\right]\\ &=2b_t^2\sum_{i=1}^{d} (\lambda_i-1)\|Ju_i\|_2^2, \end{aligned}$$ where $\lambda_i$ and $u_i$ are the eigenvalues and corresponding eigenvectors of the data covariance matrix. This expression shows that the effect of covariance mismatch on output diversity depends not only on the covariance eigenvalues, but also on how the denoising network responds to the corresponding directions. In particular, even when some data-covariance eigenvalues exceed one, the resulting output-diversity difference is not necessarily positive. Therefore, although the reviewer correctly identifies a directional covariance mismatch at initialization, this does not imply that MSIS necessarily reduces the diversity of final generated images. ## Empirical Evidence: Covariance Ablations and Direct Prompt-wise Diversity Evaluation We further investigate this concern through two complementary sets of experiments. **(a) Recall Comparison across Gaussian, Mean, and Mean + FC Initialization.** First, we revisit our covariance ablation experiments on conditional ImageNet 256 × 256 using DDIM at NFE = 6, 8, and 10. We compare three configurations: the standard Gaussian baseline, mean-based initialization (Mean), and mean initialization with full covariance matching (Mean + FC). For both Mean and Mean + FC, we use a training-data fraction of $\rho=0.01$. The complete comparison is presented below. |NFE|Initialization|FID ↓|sFID ↓|IS ↑|Precision ↑|Recall ↑| | -----| ----------------| --------: | --------: | ---------: | -------------: | ----------: | |6|Baseline|20.3680|23.5155|127.7730|0.5895|0.4330| |6|Mean|9.10|6.76|180.33|0.718|0.503| |6|Mean + FC|7.98|8.09|179.60|0.712|**0.536**| |8|Baseline|10.0229|14.3662|179.8244|0.7087|0.4642| |8|Mean|5.56|5.68|211.15|0.770|0.517| |8|Mean + FC|5.09|6.79|209.56|0.760|**0.545**| |10|Baseline|6.5124|11.3127|209.7852|0.7631|0.4784| |10|Mean|4.22|5.57|227.75|0.796|0.520| |10|Mean + FC|4.11|6.78|224.92|0.785|**0.542**| The results reveal a consistent trend across all three sampling budgets. Introducing mean-based initialization substantially improves recall over the Gaussian baseline. Furthermore, incorporating full covariance information provides additional recall improvements over mean-only initialization. Specifically, recall increases from 0.4330 to 0.503 and further to 0.536 at NFE = 6. At NFE = 8, it increases from 0.4642 to 0.517 and further to 0.545. Similarly, at NFE = 10, it increases from 0.4784 to 0.520 and further to 0.542. These results provide empirical evidence that incorporating data-dependent statistics into the initialization does not necessarily sacrifice generation coverage. In particular, full-covariance matching can further improve recall over mean-only initialization in our evaluated settings.
Reviewer_vda3OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Mean-Shifted Intermediate Sampling (MSIS), a training-free initialization strategy for few-step diffusion generation. It starts sampling from a high-noise intermediate timestep using Gaussian noise shifted by an offline estimate of the data mean. A cumulant analysis motivates the mean correction near the Gaussian endpoint, and experiments demonstrate improved sample quality across several datasets and samplers.

Critical strengths

The method is simple, requires no additional model training, and introduces negligible inference overhead. Besides, the cumulant analysis provides an intuitive motivation for prioritizing mean correction in the high-noise regime. Experiments cover multiple datasets, architectures, and samplers, with substantial improvements in several low-NFE settings.

Critical weaknesses
  1. Theorem 1 establishes an asymptotic hierarchy of data-dependent cumulants near the Gaussian endpoint, but does not provide a distributional error bound for the proposed initialization or explain how initialization error propagates through few-step sampling.
  2. The implemented initialization uses unit covariance rather than at2Ia_t^2I. Its covariance discrepancy is (1−at2)I−bt2Σ1(c)(1-a_t^2)I-b_t^2\Sigma_1(c), which is not generally second-order under the assumptions of Eq. (2).
  3. This paper does not specify how the fitting targets were obtained, the search space, selection metric, tuning data, or number of evaluations.
  4. The reported 2.22-minute preprocessing time covers mean estimation, while its batch size and timing boundaries are also unspecified. The results do not directly quantify wall-clock speedups. A complete assessment should include preprocessing and starting-point selection costs, as well as inference latency or throughput at comparable output quality. The manuscript also lacks a matched ImageNet setup-cost comparison with competing methods.
  5. Some abbreviations are not expanded at their first occurrence, formula references contain duplicated wording such as “Eq. equation,” and the hardware is inconsistently described as RTX 3090 and RTX 3090 Ti.
Questions for authors
  1. Can you provide an error bound, under a clearly specified metric, for the actual initialization btμ^1(c)+Zb_t\hat\mu_1(c)+Z? How does this error depend on the starting noise level, dimension, and mean-estimation sample size? Which evaluated paths satisfy at2=1+O(bt2)a_t^2=1+O(b_t^2)? Please also clarify whether the theoretical scope excludes flow models with non-independent endpoint couplings.
  2. How were the starting points and the linear selection rule obtained? Please report the search procedure, tuning split, selection criterion, and total evaluation cost. For the 2.22-minute mean-estimation measurement, what were the batch size and timing boundaries, and did the measurement include data loading, preprocessing, and VAE encoding?
  3. What exact distributions were used in the diagonal-variance and full-covariance ablations? How were the covariance matrices estimated and regularized? Could you additionally compare btμ^1(c)+atZb_t\hat\mu_1(c)+a_tZ against btμ^1(c)+Zb_t\hat\mu_1(c)+Z, holding the starting point and all other settings fixed?
  4. Please report end-to-end latency or throughput at matched quality, including a separate accounting of mean estimation and starting-point tuning.
  5. To investigate the proposed explanation for the CelebA-HQ results, could you provide direct covariance diagnostics or an additional evaluation on FFHQ?
13 comments
AuthorsOpenReview ↗

Response to Q1-1: Empirical Distributional Error and Theoretical Scope

Thank you for raising these important questions. We agree that the cumulant hierarchy established by Theorem 1 does not directly provide a quantitative distributional error bound for the practical initialization. To address this concern, we provide an additional Wasserstein-2 error decomposition that separates the intrinsic approximation error from the finite-sample mean-estimation error. Under the i.i.d. sampling assumption, we derive an explicit bound for the latter, characterizing its dependence on the starting noise level, data dimensionality, and estimation sample size.

To further assess the practical approximation quality, we conduct additional experiments using the sliced Wasserstein distance (SWD). Since each ImageNet class contains only approximately 1,300 original training images, using the same images for both mean estimation and empirical distribution evaluation could introduce an in-sample data-reuse effect, potentially leading to an optimistically biased (underestimated) distributional discrepancy. To avoid this potential bias, we instead use intermediate states obtained from a separately generated 50-step DDIM reverse trajectory as the reference distributions. We then measure the SWD between the MSIS initialization and the corresponding DDIM-induced intermediate distributions. These experiments investigate the effects of the starting timestep, the amount of data used for mean estimation, and the evaluation space.

We also clarify the theoretical scope of the coefficient-scaling condition and the conditional independence assumption between the Gaussian base variable and the data variable.

In the revised manuscript, we will carefully check the notation and terminology throughout to ensure consistency. In particular, we will standardize the notation for the estimated data mean and the timestep-dependent Gaussian-noise scaling coefficient, correct equation references and terminology, and ensure that abbreviations and hardware specifications are consistently described.

Distributional error bound

We thank the reviewer for pointing out the distinction between the asymptotic cumulant analysis and a quantitative distributional error bound. We agree that Theorem 1 characterizes the orders of moment and cumulant discrepancies near the Gaussian endpoint, but does not establish a finite-time Wasserstein or total variation bound for the complete initialization distribution.

To provide a quantitative analysis of the finite-sample initialization error, we decompose it into two components: an intrinsic approximation error and a mean-estimation error.

Let Ptc=L(xt⋆∣c)P_t^c=\mathcal{L}(x_t^\star\mid c) denote the true intermediate distribution, Qtc=N(btμ1(c),Id)Q_t^c=\mathcal{N}(b_t\mu_1(c),I_d) the ideal mean-shifted Gaussian distribution using the population mean, and Q^tc=N(btμ^1(c),Id)\hat Q_t^c=\mathcal{N}(b_t\hat\mu_1(c),I_d) the actual initialization using the estimated mean. Using the Wasserstein-2 distance as the distributional metric, the triangle inequality gives

W2(Ptc,Q^tc)≤W2(Ptc,Qtc)+W2(Qtc,Q^tc).W_2(P_t^c,\hat Q_t^c) \leq W_2(P_t^c,Q_t^c) + W_2(Q_t^c,\hat Q_t^c).

Since the latter two Gaussian distributions have identical covariance matrices, their Wasserstein-2 distance is determined solely by the difference in their means:

W2(Qtc,Q^tc)=∣bt∣∥μ^1(c)−μ1(c)∥2.W_2(Q_t^c,\hat Q_t^c) = |b_t|\|\hat\mu_1(c)-\mu_1(c)\|_2.

Assuming that μ^1(c)\hat\mu_1(c) is estimated from ncn_c independent and identically distributed (i.i.d.) samples with finite covariance matrix Σ1(c)\Sigma_1(c), we obtain

EDc[∥μ^1(c)−μ1(c)∥22]=tr⁡(Σ1(c))nc,\mathbb{E}_{\mathcal{D}_c} \left[ \|\hat\mu_1(c)-\mu_1(c)\|_2^2 \right] = \frac{\operatorname{tr}(\Sigma_1(c))}{n_c},

where Dc\mathcal{D}_c denotes the random estimation dataset.

Consequently,

EDc[W2(Qtc,Q^tc)]≤∣bt∣tr⁡(Σ1(c))nc,\mathbb{E}_{\mathcal{D}_c} \left[ W_2(Q_t^c,\hat Q_t^c) \right] \leq |b_t| \sqrt{ \frac{\operatorname{tr}(\Sigma_1(c))}{n_c} },

and therefore,

EDc[W2(Ptc,Q^tc)]≤Δt(c)+∣bt∣tr⁡(Σ1(c))nc,\mathbb{E}_{\mathcal{D}_c} \left[ W_2(P_t^c,\hat Q_t^c) \right] \leq \Delta_t(c) + |b_t| \sqrt{ \frac{\operatorname{tr}(\Sigma_1(c))}{n_c} },

where Δt(c)=W2(Ptc,Qtc)\Delta_t(c)=W_2(P_t^c,Q_t^c) denotes the intrinsic approximation error.

Furthermore, if ∥Σ1(c)∥op≤σc2\|\Sigma_1(c)\|_{\mathrm{op}}\leq\sigma_c^2, the mean-estimation error term can be further bounded by

∣bt∣σcdnc.|b_t|\sigma_c\sqrt{\frac{d}{n_c}}.

This explicitly characterizes the dependence of the mean-estimation error on the starting timestep (through btb_t), data dimensionality, and estimation sample size.

Importantly, the covariance discrepancy between the ideal initialization and the true intermediate distribution is given by

Cov⁡(Qtc)−Cov⁡(Ptc)=(1−at2)Id−bt2Σ1(c).\operatorname{Cov}(Q_t^c) - \operatorname{Cov}(P_t^c) = (1-a_t^2)I_d-b_t^2\Sigma_1(c).

For paths satisfying the additional scaling condition at2=1+O(bt2)a_t^2=1+O(b_t^2), this discrepancy is O(bt2)O(b_t^2) in operator norm, provided that ∥Σ1(c)∥op\|\Sigma_1(c)\|_{\mathrm{op}} is bounded. Meanwhile, as established in Theorem 1, for each fixed order 3≤r≤R3\leq r\leq R, the higher-order cumulant discrepancies remain O(btr)O(b_t^r).

AuthorsOpenReview ↗

Response to Q1-2: Empirical Distributional Error and Theoretical Scope

We emphasize that these results quantify the finite-sample mean-estimation error and the moment-level approximation behavior, rather than providing a complete finite-time Wasserstein bound for Δt(c)\Delta_t(c). Establishing such a bound would require additional distributional regularity assumptions beyond those considered in our current analysis. We therefore complement the theoretical results with empirical SWD measurements to assess the approximation quality at practical starting timesteps.

Experimental setup and distributional measurements

Our practical initialization is

XtMSIS=btμ^1(c)+Z,Z∼N(0,I).X_t^{\mathrm{MSIS}} = b_t\hat\mu_1(c)+Z, \qquad Z\sim\mathcal N(0,I).

We directly compare its distribution against the actual intermediate distribution obtained from the DDIM reverse trajectory, rather than relying on an idealized analytical forward marginal.

For the DiT-XL/2-256 experiment, we use 1,000 classes and generate 650 reverse samples per class, resulting in 650,000 evaluation samples at each timestep. The reverse trajectory is a 50-step DDIM trajectory starting from t=999t=999. We evaluate seven requested timesteps, whose corresponding native DDIM timesteps are 899, 859, 799, 759, 699, 659, and 599, respectively.

The class-conditional mean and covariance statistics are estimated independently from real ImageNet data latents corresponding to x1x_1. We consider three statistics fractions:

ρ∈{0.01,0.1,1.0}.\rho\in\{0.01,0.1,1.0\}.

We evaluate both a low-frequency subspace, retaining the 81 lowest-frequency positions out of 1,024 per channel, and the full latent space. Since the early stages of the denoising process primarily establish coarse image structures and global contours, which are mainly captured by low-frequency components, we focus on evaluating the improvement in the low-frequency subspace. For completeness, we also report the results over the full latent space. All SWD measurements use 256 random projection directions.

At each timestep, we compare the empirical distribution of the actual DDIM reverse-sampling states, denoted by P^treverse\hat P_t^{\mathrm{reverse}}, against the following five reference distributions:

Q_0 &= \mathcal{N}(0,I),\\ Q_1 &= \mathcal{N}(0,a_t^2I),\\ Q_2 &= \mathcal{N}(b_t\hat{\mu}_1(c),I),\\ Q_3 &= \mathcal{N}(b_t\hat{\mu}_1(c),a_t^2I),\\ Q_4 &: \text{mean-shifted Gaussian with estimated}\\ &\qquad \text{projection-dependent covariance}. \end{aligned}$$ We then compute the SWD between the actual reverse-sampling distribution and each reference distribution: $$D_i(t)= \operatorname{SWD}_{256} \left( \hat P_t^{\mathrm{reverse}},Q_i \right), \qquad i=0,\ldots,4.$$ Here, $D_2$ corresponds exactly to our practical MSIS initialization. More precisely, each reported distance is the SWD between the actual reverse-sampling distribution and the corresponding reference distribution. We report the following improvements: $$I_1=\frac{D_0-D_1}{D_0}\times100\%, \qquad I_2=\frac{D_0-D_2}{D_0}\times100\%.$$ For the two covariance-related comparisons, we use a common denominator: $$I_3=\frac{D_2-D_3}{D_0}\times100\ \mathrm{pp},$$ $$I_4=\frac{D_2-D_4}{D_0}\times100\ \mathrm{pp}.$$ Positive values indicate SWD reductions, whereas negative values indicate increased SWD. ## Complete SWD results **Low-frequency subspace,** $\rho=0.01$​. |Requested / actual t|D0|D1|D1 vs D0|D2|D2 vs D0|D3|D3 vs D2|D4|D4 vs D2| | ----------------------| --------: | --------: | ---------: | --------: | ---------: | --------: | ---------: | --------: | ---------: | |900 / 899|0.07209|0.07209|+0.00%|0.07135|+1.03%|0.07134|+0.00 pp|0.07136|-0.01 pp| |850 / 859|0.07441|0.07440|+0.00%|0.07236|+2.74%|0.07236|+0.00 pp|0.07237|-0.01 pp| |800 / 799|0.08148|0.08149|-0.01%|0.07572|+7.07%|0.07574|-0.02 pp|0.07569|+0.05 pp| |750 / 759|0.08990|0.08996|-0.06%|0.07996|+11.06%|0.08002|-0.07 pp|0.07979|+0.19 pp| |700 / 699|0.11036|0.11067|-0.28%|0.09139|+17.19%|0.09175|-0.33 pp|0.09046|+0.84 pp| |650 / 659|0.13146|0.13228|-0.62%|0.10447|+20.53%|0.10546|-0.75 pp|0.10210|+1.80 pp| |600 / 599|0.17736|0.18023|-1.62%|0.13592|+23.36%|0.13946|-2.00 pp|0.12811|+4.40 pp| **Full latent space,** $\rho=0.01$. |Requested / actual t|D0|D1|D1 vs D0|D2|D2 vs D0|D3|D3 vs D2|D4|D4 vs D2| | ----------------------| --------: | --------: | ---------: | --------: | ---------: | --------: | ---------: | --------: | ---------: | |900 / 899|0.07083|0.07083|+0.01%|0.07081|+0.03%|0.07081|+0.01 pp|0.07081|+0.00 pp| |850 / 859|0.07114|0.07113|+0.01%|0.07102|+0.16%|0.07102|+0.01 pp|0.07102|+0.00 pp| |800 / 799|0.07172|0.07171|+0.02%|0.07136|+0.51%|0.07134|+0.02 pp|0.07135|+0.01 pp| |750 / 759|0.07251|0.07249|+0.03%|0.07180|+0.98%|0.07178|+0.03 pp|0.07179|+0.01 pp| |700 / 699|0.07494|0.07491|+0.04%|0.07330|+2.19%|0.07327|+0.04 pp|0.07328|+0.03 pp| |650 / 659|0.07760|0.07758|+0.03%|0.07501|+3.35%|0.07498|+0.03 pp|0.07497|+0.05 pp| |600 / 599|0.08412|0.08422|-0.13%|0.07935|+5.67%|0.07946|-0.14 pp|0.07927|+0.09 pp|
AuthorsOpenReview ↗

Response to Q1-3: Empirical Distributional Error and Theoretical Scope

Low-frequency subspace, ρ=0.1\rho=0.1.

Requested / actual tD0D1D1 vs D0D2D2 vs D0D3D3 vs D2D4D4 vs D2
900 / 8990.072090.07209+0.00%0.07094+1.60%0.07094+0.00 pp0.07095-0.01 pp
850 / 8590.074410.07440+0.00%0.07152+3.87%0.07152+0.00 pp0.07153-0.01 pp
800 / 7990.081480.08149-0.01%0.07356+9.73%0.07357-0.02 pp0.07351+0.05 pp
750 / 7590.089900.08996-0.06%0.07615+15.29%0.07622-0.07 pp0.07595+0.23 pp
700 / 6990.110360.11067-0.28%0.08346+24.38%0.08385-0.36 pp0.08231+1.04 pp
650 / 6590.131460.13228-0.62%0.09246+29.67%0.09356-0.83 pp0.08942+2.31 pp
600 / 5990.177360.18023-1.62%0.11598+34.61%0.12008-2.31 pp0.10550+5.91 pp

Full latent space, ρ=0.1\rho=0.1.

Requested / actual tD0D1D1 vs D0D2D2 vs D0D3D3 vs D2D4D4 vs D2
900 / 8990.070830.07083+0.01%0.07074+0.13%0.07074+0.01 pp0.07074+0.00 pp
850 / 8590.071140.07113+0.01%0.07089+0.35%0.07088+0.01 pp0.07088+0.00 pp
800 / 7990.071720.07171+0.02%0.07099+1.01%0.07098+0.02 pp0.07099+0.01 pp
750 / 7590.072510.07249+0.03%0.07113+1.91%0.07110+0.03 pp0.07112+0.01 pp
700 / 6990.074940.07491+0.04%0.07177+4.23%0.07174+0.04 pp0.07175+0.03 pp
650 / 6590.077600.07758+0.03%0.07239+6.72%0.07237+0.03 pp0.07234+0.06 pp
600 / 5990.084120.08422-0.13%0.07402+12.01%0.07414-0.14 pp0.07389+0.15 pp

Low-frequency subspace, ρ=1.0\rho=1.0.

Requested / actual tD0D1D1 vs D0D2D2 vs D0D3D3 vs D2D4D4 vs D2
900 / 8990.072090.07209+0.00%0.07091+1.64%0.07090+0.00 pp0.07092-0.01 pp
850 / 8590.074410.07440+0.00%0.07146+3.96%0.07145+0.00 pp0.07146-0.01 pp
800 / 7990.081480.08149-0.01%0.07338+9.94%0.07339-0.02 pp0.07334+0.06 pp
750 / 7590.089900.08996-0.06%0.07585+15.63%0.07591-0.07 pp0.07564+0.24 pp
700 / 6990.110360.11067-0.28%0.08276+25.01%0.08315-0.36 pp0.08158+1.07 pp
650 / 6590.131460.13228-0.62%0.09136+30.50%0.09247-0.84 pp0.08824+2.37 pp
600 / 5990.177360.18023-1.62%0.11409+35.68%0.11825-2.35 pp0.10331+6.08 pp

Full latent space, ρ=1.0\rho=1.0.

Requested / actual tD0D1D1 vs D0D2D2 vs D0D3D3 vs D2D4D4 vs D2
900 / 8990.070830.07083+0.01%0.07074+0.13%0.07074+0.01 pp0.07074+0.00 pp
850 / 8590.071140.07113+0.01%0.07088+0.37%0.07087+0.01 pp0.07087+0.00 pp
800 / 7990.071720.07171+0.02%0.07096+1.05%0.07095+0.02 pp0.07096+0.01 pp
750 / 7590.072510.07249+0.03%0.07108+1.97%0.07106+0.03 pp0.07107+0.01 pp
700 / 6990.074940.07491+0.04%0.07163+4.41%0.07160+0.04 pp0.07161+0.03 pp
650 / 6590.077600.07758+0.03%0.07214+7.04%0.07212+0.03 pp0.07209+0.06 pp
600 / 5990.084120.08422-0.13%0.07353+12.59%0.07365-0.15 pp0.07340+0.16 pp

The above results reveal four important observations.

Observation 1: Scaling only the Gaussian variance does not provide a consistent improvement. Comparing D1D_1 against D0D_0, we find that changing the Gaussian variance without correcting the mean provides almost no benefit. In the low-frequency subspace, the improvement is approximately zero at the earliest evaluated timesteps and becomes negative as sampling proceeds. At t=599t=599, it increases SWD by 1.62%. In the full latent space, the effect is also very small and becomes negative at t=599t=599 (-0.13%). This indicates that simply rescaling the Gaussian noise is insufficient to capture the data-dependent structure of the intermediate distribution.

Observation 2: The benefit of mean correction increases with the statistics fraction. Comparing D2D_2 against D0D_0, we observe a consistent improvement as the statistics fraction increases from 0.01 to 0.1 and 1.0. At t=599t=599, the low-frequency SWD reductions are 23.36%, 34.61%, and 35.68%, respectively. In the full latent space, the corresponding reductions are 5.67%, 12.01%, and 12.59%. The improvement is therefore strongly affected by the number of samples used to estimate the mean, although the marginal benefit becomes smaller when increasing the statistics fraction from 0.1 to 1.0.

Reviewer_XhtNOpenReview ↗
Rating
3
Confidence
3(+1)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Mean-Shifted Intermediate Sampling (MSIS), a training-free method for accelerating few-step diffusion generation. By approximating the intermediate distribution with mean-shifted Gaussian noise, MSIS shortens the reverse sampling trajectory while introducing negligible inference overhead.

Critical strengths
  1. Simple and practical approach. MSIS requires no additional model training and can be integrated with different pretrained samplers and timestep schedules with minimal modifications.

  2. Strong experimental results. The method achieves substantial FID improvements across several datasets and sampling algorithms, particularly in the few-step regime. The results on conditional ImageNet 256×256 are especially encouraging.

  3. Useful theoretical motivation and ablations. The cumulant analysis provides an explanation for the mean-shifted approximation. The component, data-fraction, and covariance ablations further support the effectiveness and efficiency of the proposed design.

Critical weaknesses

W1: Starting-timestep selection. The performance of MSIS is sensitive to the starting timestep, as shown in Figure 2. Although Appendix A describes a linear fitting strategy, the calibration procedure and its associated computational cost could be explained more clearly. This would help readers better understand the practical overhead of applying MSIS to new models or schedules.

W2: Connection between theory and practice. The theoretical analysis establishes the dominance of the mean shift asymptotically near the Gaussian endpoint. However, it remains unclear how closely this theoretical characterization applies to the finite starting timesteps used in the experiments. Additional empirical evidence on the relative contributions of the mean shift and higher-order distributional differences at these timesteps would strengthen the theoretical justification.

Questions for authors

Q1: Could the authors provide more details on how the starting timesteps are determined, particularly the calibration procedure and any additional computational overhead?

Q2: Could the authors further clarify how well the asymptotic analysis describes the intermediate distributions at the starting timesteps used in practice?

Q3: Could the authors provide end-to-end sampling latency or quality–latency comparisons to further demonstrate the practical acceleration benefits of MSIS?

7 comments
AuthorsOpenReview ↗

Response to Q1-1: Practical Configuration and Efficient Calibration of MSIS

We thank the reviewer for this helpful question.

MSIS is a training-free method. It has only two method-specific settings: the data fraction ρ\rho used for mean estimation and the starting timestep tst_s. Table 2(d) shows that MSIS is insensitive to ρ\rho. Using only 1% of the training data gives results close to those obtained with 50% or 100% of the data. tst_s is the only setting that requires careful calibration.

This is an important practical advantage of MSIS. Compared with methods that introduce additional learned components, MSIS avoids the need to configure and tune a separate training process, such as its optimizer and training schedule. The mean statistics are computed only once offline. They can be reused across different samplers, timestep schedules, and models that share the same input space. For any candidate starting timestep, we only need to scale the precomputed mean using the corresponding forward coefficient. We do not need to recompute the statistics.

complete calibration procedure

We use conditional ImageNet 256×256 with DDIM as an example to explain the complete calibration procedure and its computational cost. All measurements below are conducted on a single RTX 3090.

1. Mean estimation.

Following Table 2(c), we use ρ=0.01\rho=0.01 to estimate the class-conditional means. This takes ​2 min 13.20 s.

2. KID reference feature extraction.

We randomly select 10K training images as a proxy reference set for KID evaluation. We extract and cache their features once. This takes ​2 min 14.81 s. The same reference features are reused for all candidate starting points.

3. Coarse-to-fine discrete search.

We use a multi-stage coarse-to-fine discrete search to determine the starting timestep.

There are two main sources of error when moving the starting point from the noise end toward the data end. The first is the discretization error caused by large reverse steps. The second is the distribution mismatch introduced by approximating the intermediate distribution with mean-shifted Gaussian noise.

These two errors have different effects. At an earlier starting point, the reverse trajectory is longer, so discretization error is more dominant. At a later starting point, the mean-shifted Gaussian approximation becomes less accurate, so distribution mismatch becomes more important. This trade-off leads to a U-shaped performance curve.

Our search procedure is as follows:

  • Round 1: We first evaluate starting timesteps [200, 400, 600, 800], with an interval of 200. The KID values at 200 and 800 are relatively high, while those at 400 and 600 are lower. We therefore narrow the search to [400, 800].
  • Round 2: We reduce the interval to 100 by adding 500 and 700. The promising region is further narrowed to [500, 700].
  • Round 3: We reduce the interval to 50 by adding 550 and 650. We stop when the interval reaches the predefined resolution of 50. The selected starting timestep is 600.

This process takes three rounds and requires ​8 evaluations for NFE=6.

4. Search for the maximum NFE and linear fitting.

We first search for the starting timestep at the minimum NFE (6). We then search for the maximum NFE (10). The lower bound of the second search is set to the selected starting timestep for NFE=6. This follows the general observation that a larger NFE budget usually allows sampling to start at an equal or later point along the trajectory.

The NFE=10 search requires 6 additional evaluations. Therefore, the complete search requires ​14 evaluations.

We estimate the search time using the average end-to-end evaluation time measured in our 1K-sample runs. The average time is 84.8 s per evaluation for NFE=6 and 106.2 s for NFE=10. The total search time is:

  • NFE=6: 8 × 84.8 s = 678.4 s (11 min 18.4 s).
  • NFE=10: 6 × 106.2 s = 637.2 s (10 min 37.2 s).

The total search time is approximately ​21 min 55.6 s.

After obtaining the two starting points, (NFE=6, ts=600t_s=600) and (NFE=10, ts=700t_s=700), we fit a linear rule between NFE and the starting timestep:

ts=25×NFE+450.t_s = 25 \times \mathrm{NFE} + 450.

This gives ts=650t_s=650 for NFE=8. In this way, we can determine the starting timestep for any NFE within the calibrated range without performing another search. The selected starting points are then reused across numerical samplers under the same model and schedule.

The complete time breakdown is shown below.

Calibration stageTimePercentage
One-time preparation (mean estimation + KID feature extraction)4 min 28.01 s16.9%
Starting-point search (14 evaluations)21 min 55.60 s83.1%

The total calibration time is ​26 min 23.61 s.

AuthorsOpenReview ↗

Response to Q1-2: Practical Configuration and Efficient Calibration of MSIS

We emphasize that selecting a suitable starting point is a general consideration in intermediate sampling methods. It requires balancing the benefit of shortening the reverse trajectory against the distribution mismatch at the chosen starting point. This is not specific to MSIS. Our method makes this process lightweight because it only requires evaluating candidate starting points and does not involve additional model training.

For reference, AC-Sampler reports less than two hours of training on CIFAR-10 using a single RTX 3090. It does not report the corresponding training cost on ImageNet. Although these are not directly comparable settings, our complete calibration on ImageNet 256×256 takes only about 26 minutes on a single RTX 3090, including the starting-point search. No additional model training is required.

Reliability of KID as a search metric

The standard evaluation protocol uses 50K generated samples, which is expensive for repeated candidate evaluations. We therefore use 1K generated samples per candidate and KID as a proxy metric. KID is suitable for comparisons with a limited number of samples.

We further compare KID with FID under DDIM with NFE=6. The results are shown below. KID and the 1K-sample FID use 1K generated samples, while the 50K-sample FID follows the standard evaluation protocol.

Starting timestepKID ↓ (1K samples, mean ± std)FID ↓ (1K samples)FID ↓ (50K samples)
5000.006897 ± 0.00049452.4012.94
5500.005240 ± 0.00038749.6010.27
6000.005024 ± 0.00035848.599.10
6500.005090 ± 0.00034048.539.91
7000.005867 ± 0.00036950.6210.30

The results show that KID and the standard 50K-sample FID both reach their minimum at a starting timestep of 600. The 1K-sample FID reaches its minimum at 650, but the difference between 600 and 650 is only 0.06 (approximately 0.12%). Thus, the two settings give nearly identical results under the 1K-sample evaluation.

All three metrics show a consistent overall trend: generation quality first improves and then degrades as the starting timestep increases. In particular, KID correctly identifies the optimal starting timestep under the standard 50K-sample FID evaluation. These results support the reliability of KID as a proxy metric for efficient starting-point calibration. They also show that using KID with only 1K generated samples can substantially reduce the search cost without changing the selected starting timestep.

We will incorporate the complete calibration procedure, the corresponding time breakdown, and the KID–FID consistency results into the revised manuscript to clarify the practical configuration of MSIS and facilitate reproducibility.

AuthorsOpenReview ↗

Response to Q2-1: Empirical Validity of the Asymptotic Analysis at Practical Starting Timesteps

Thank you for raising this important question. We agree that the asymptotic analysis established in Theorem 1 alone does not directly quantify how accurately the mean-shifted Gaussian approximation describes the actual intermediate distributions at the finite starting timesteps used in practice. To further investigate the applicability of our theoretical analysis in practical sampling scenarios, we conduct additional distributional measurements that directly compare the practical MSIS initialization with the actual intermediate distributions obtained along the DDIM reverse sampling trajectory.

Experimental setup and distributional measurements

During the early stages of reverse denoising, diffusion models primarily establish coarse image structures and global contours, which are mainly captured by low-frequency components. Therefore, we focus on evaluating distribution matching in the low-frequency subspace. Meanwhile, to provide a comprehensive assessment of the overall distributional improvement introduced by mean correction, we also report results in the full latent space. Specifically, the low-frequency subspace retains the 81 lowest-frequency spatial positions out of 1,024 positions per channel.

We conduct experiments on DiT-XL/2-256 across 1,000 ImageNet classes, generating 650 reverse-sampling samples per class and resulting in 650,000 samples at each evaluation timestep. We use a 50-step DDIM reverse trajectory starting from timestep 999 and evaluate the intermediate distributions at seven timesteps, corresponding to native DDIM timesteps 899, 859, 799, 759, 699, 659, and 599. The class-conditional means and covariance statistics are independently estimated from real ImageNet data, using 1% of the training data, consistent with our main experimental setting. All SWD measurements use 256 random projection directions.

To investigate the contributions of different distributional corrections, we compare the empirical distribution obtained from the actual DDIM reverse trajectory, denoted by P^treverse\hat P_t^{\mathrm{reverse}}, against the following five reference distributions:

Q_0 &= \mathcal{N}(0,I),\\ Q_1 &= \mathcal{N}(0,a_t^2I),\\ Q_2 &= \mathcal{N}(b_t\hat{\mu}_1(c),I),\\ Q_3 &= \mathcal{N}(b_t\hat{\mu}_1(c),a_t^2I),\\ Q_4 &: \text{mean-shifted Gaussian with estimated}\\ &\qquad \text{projection-dependent covariance}. \end{aligned}$$ Here, $Q_2$ corresponds exactly to the practical MSIS initialization adopted in our method. We calculate the SWD between the actual intermediate distribution and each of the five reference distributions: $$D_i(t)= \operatorname{SWD}_{256} \left( \hat P_t^{\mathrm{reverse}},Q_i \right), \quad i=0,\ldots,4.$$ ## **Complete SWD results** We first report the complete results in the low-frequency subspace. **Low-frequency subspace,** $p=0.01$ |Requested / actual timestep|$D_0$|$D_1$|$D_1$ vs. $D_0$|$D_2$|$D_2$ vs. $D_0$|$D_3$|$D_3$ vs. $D_2$|$D_4$|$D_4$ vs. $D_2$| | -----------------------------| ---------| ---------| --------| ---------| ---------| ---------| ----------| ---------| ----------| |900 / 899|0.07209|0.07209|+0.00%|0.07135|+1.03%|0.07134|+0.00 pp|0.07136|-0.01 pp| |850 / 859|0.07441|0.07440|+0.00%|0.07236|+2.74%|0.07236|+0.00 pp|0.07237|-0.01 pp| |800 / 799|0.08148|0.08149|-0.01%|0.07572|+7.07%|0.07574|-0.02 pp|0.07569|+0.05 pp| |750 / 759|0.08990|0.08996|-0.06%|0.07996|+11.06%|0.08002|-0.07 pp|0.07979|+0.19 pp| |700 / 699|0.11036|0.11067|-0.28%|0.09139|+17.19%|0.09175|-0.33 pp|0.09046|+0.84 pp| |650 / 659|0.13146|0.13228|-0.62%|0.10447|+20.53%|0.10546|-0.75 pp|0.10210|+1.80 pp| |600 / 599|0.17736|0.18023|-1.62%|0.13592|+23.36%|0.13946|-2.00 pp|0.12811|+4.40 pp| We next report the results in the full latent space. **Full latent space,** **$\rho=0.01$** |Requested / actual timestep|$D_0$|$D_1$|$D_1$ vs. $D_0$|$D_2$|$D_2$ vs. $D_0$|$D_3$|$D_3$ vs. $D_2$|$D_4$|$D_4$ vs. $D_2$| | -----------------------------| ---------| ---------| --------| ---------| --------| ---------| ----------| ---------| ----------| |900 / 899|0.07083|0.07083|+0.01%|0.07081|+0.03%|0.07081|+0.01 pp|0.07081|+0.00 pp| |850 / 859|0.07114|0.07113|+0.01%|0.07102|+0.16%|0.07102|+0.01 pp|0.07102|+0.00 pp| |800 / 799|0.07172|0.07171|+0.02%|0.07136|+0.51%|0.07134|+0.02 pp|0.07135|+0.01 pp| |750 / 759|0.07251|0.07249|+0.03%|0.07180|+0.98%|0.07178|+0.03 pp|0.07179|+0.01 pp| |700 / 699|0.07494|0.07491|+0.04%|0.07330|+2.19%|0.07327|+0.04 pp|0.07328|+0.03 pp| |650 / 659|0.07760|0.07758|+0.03%|0.07501|+3.35%|0.07498|+0.03 pp|0.07497|+0.05 pp| |600 / 599|0.08412|0.08422|-0.13%|0.07935|+5.67%|0.07946|-0.14 pp|0.07927|+0.09 pp|

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.