Computer Vision and Pattern Recognition · arXiv:2611.17194 v1 · submitted 12 Nov 2026 · 68 references
eDMD: Endpoint Distribution Matching for Few-Step Diffusion Distillation
The paper says
“Distribution matching distillation (DMD) has emerged as a leading paradigm for distilling diffusion models into few-step students, where an auxiliary fake score model tracks the student's distribution and the discrepancy between the teacher and fake predictions drives the student toward the teacher.”
more of the abstract
In multi-step distillation, although only the student's final output is delivered, existing methods train the fake model on intermediate states or across all denoising steps. Consequently, the student's update signal becomes entangled with intermediate distributions, which can distort the matching direction and remain nonzero even after the student's final outputs already match the teacher distribution. To resolve this mismatch, we propose Endpoint Distribution Matching Distillation (eDMD), which trains a single shared fake model exclusively on the student's final outputs and uses it to correct every denoising step. Our theoretical analysis proves that this endpoint reference guarantees that the distribution matching term vanishes if and only if the student's final output distribution matches the teacher's, regardless of intermediate distributions. Beyond the fake reference, we further strengthen teacher supervision through a precise teacher (PT) that replaces single-step guided predictions with short multi-step trajectories on the final noise interval, paired with a calibrated query density for intermediate steps. Experiments on SD3-Medium, FLUX.2-klein (4B and 9B), and Qwen-Image-2.1 demonstrate that eDMD consistently outperforms existing distribution matching methods for 4-step generation.
from the abstract
Pith says
Machine review · deepseek-v4-flash · 13 Nov 2026 · what was read was not recorded · not checked by a person
It can be wrong. Challenge this review or reply to it.
The passes disagreed. Reader: holds up. Skeptic: agrees with the reader. Referee: major revision.
A clean theoretical fix for a known DMD issue, backed by strong experiments; the practical benefit is real even if the theoretical guarantee is conditional on exact denoisers. This line is wrong
4 major objections · 9 minor
How often objections from reviews like this one are right: not yet measured
Not checked: figures, code, what the 68 references say.
The paper says, at §3.2, Proposition 1
The review did not quote this passage.
Pith says: Objection 1 of 4 · major
The theoretical guarantee assumes exact conditional denoisers under a shared forward noising process. The paper acknowledges this (b85, b290) and provides empirical evidence that the endpoint reference works well. However, the central claim that eDMD 'eliminates the reference residual' in practice depends on how well the endpoint fake approximates D_q under neural approximation error and finite data.
more of objection 1
The paper does not analyze how approximation error in the fake model propagates into the correction signal C = A - F. In particular, if F is a biased estimate of D_q, the residual R = D_q - F may not vanish even when ν = q, and the correction may not vanish when q = p. The ablations (Tab. 3, w/o E) show that removing the endpoint reference causes the largest degradation, but this does not isolate the effect of fake model approximation error from the effect of the reference distribution mismatch. The concern is that the practical benefit of eDMD may be partly due to other components (PT, calibrated query density) or to the endpoint reference providing a better-conditioned learning signal, rather than to the elimination of the reference residual per se. The paper's own analysis of the Gaussian case (Lemma 4, App. A.5) shows that even with an exact endpoint fake, the student's equilibrium variance can differ from the teacher's due to the interplay of PT and query density, suggesting that the endpoint reference alone does not guarantee exact matching in practice. The authors should discuss this limitation more explicitly and ideally provide an analysis of how approximation error in the fake model affects the correction signal.
The paper says, at §5.3, Table 3
The review did not quote this passage.
Pith says: Objection 2 of 4 · major
The ablation study removes one component at a time, but the 'w/o E' ablation replaces the endpoint fake with a stage-based fake (TDM's weighting). This changes both the reference distribution and the fake training objective. It would be more informative to also ablate by training the fake on final outputs but with a different loss (e.g., the same loss as …
more of objection 2
… w/o E but on final outputs only) to isolate the effect of the reference distribution from the effect of the training objective. The paper's claim that 'removing E causes the largest overall degradation' is supported, but the ablation conflates multiple changes.
The paper says, at §4.2, Precise Teacher
The review did not quote this passage.
Pith says: Objection 3 of 4 · major
The PT component uses short multi-step integration for low-noise queries. The paper reports that extending PT to all queries causes training collapse (App. C). This suggests that the benefit of PT is specific to low-noise levels and that the multi-step integration may introduce bias at higher noise levels.
more of objection 3
The paper should discuss why PT is harmful at higher noise levels and how the threshold σ_{K-1} was chosen. Is there a principled way to set this threshold, or is it a hyperparameter that needs tuning per teacher?
The paper says, at §4.3, Calibrated query density
The review did not quote this passage.
Pith says: Objection 4 of 4 · major
The calibrated query density is computed once before training using teacher samples. The paper reports that on FLUX.2-klein-4B, the calibrated density is nearly uniform and performance is similar to uniform sampling (Tab. 3). This suggests that the calibration may not be necessary for all teachers. The paper should discuss when calibration is expected to be beneficial and whether the calibration procedure could be simplified or omitted for some architectures.
9 minor objections
The paper says, at §1, Introduction
Pith says: minor objection 1 of 9
The phrase 'the fake model should represent the distribution of the samples that distillation ultimately aims to optimize' is a key insight but could be stated more precisely. Consider rephrasing to clarify that the fake model should be trained on the student's final outputs, not on intermediate predictions.
The paper says, at §3.2, Eq. (5)
Pith says: minor objection 2 of 9
The derivation of Eq. (5) for the mixture reference residual is given in App. A.3. It would be helpful to include a brief intuitive explanation in the main text of why the residual pushes final outputs toward regions where the last step's density exceeds that of earlier steps.
show 7 more
The paper says, at §4.1, Eq. (7)
Pith says: minor objection 3 of 9
The fake training loss uses a clipped SNR weight. The paper should specify the clipping range (min{max{SNR(τ),1},10}) in the main text or at least reference the appendix where it is given (App. C).
The paper says, at §5.1, Evaluation
Pith says: minor objection 4 of 9
The evaluation uses DrawBench-1000 (200 prompts × 5 seeds). The paper should report whether the 200 prompts are a standard subset or were selected specifically for this study. If a subset, how was it chosen?
The paper says, at §5.2, Table 1
Pith says: minor objection 5 of 9
The table includes 'Distilled (4 steps)*' for FLUX.2-klein models as an unranked reference. It would be helpful to note in the caption or text what method was used to produce these official checkpoints (e.g., adversarial distillation, progressive distillation, etc.).
The paper says, at §5.3, Ablations
Pith says: minor objection 6 of 9
The rerun noise analysis reports standard deviations of 0.074 on SD3-Medium and 0.11 on FLUX.2-klein-4B. The paper should clarify whether these are standard deviations of the HPSv3++ score across checkpoints or across seeds. If across seeds, the differences between ablations (e.g., w/o E vs. full) are much larger than the noise, supporting the conclusions.
The paper says, at App. A.5, Lemma 4
Pith says: minor objection 7 of 9
The Gaussian analysis is insightful but assumes a scale-only final step. The paper should note that this is a simplified setting and that the conclusions may not directly generalize to high-dimensional, non-Gaussian distributions.
The paper says, at App. C, Training settings
Pith says: minor objection 8 of 9
The paper uses LoRA for the student and full parameters for the fake on FLUX.2-klein-9B. The choice of LoRA rank (32) and alpha (64) should be justified or referenced to prior work.
The paper says, at App. H, Endpoint reference on TDM
Pith says: minor objection 9 of 9
The experiment adding the endpoint reference to TDM shows improvements, but the gains are smaller than those of full eDMD. The paper should discuss why the endpoint reference alone does not fully close the gap to eDMD, i.e., what additional benefits PT and calibrated query density provide.
What would settle it
Train eDMD on a small-scale problem (e.g., 2D synthetic or CIFAR-10) where the exact denoiser D_q can be computed analytically or via a sufficiently large model. Compare the correction C = A - F against the theoretical correction C_theory = A - D_q. If the difference ||C - C_theory|| is large relative to the norm of C_theory, then fake model approximation error is significant and the 'elimination' of the reference residual is incomplete. Alternatively, on the same small-scale problem, train a student with the endpoint reference but replace F with a Monte Carlo estimate of D_q (e.g., by averaging many samples). If the performance gap between this 'oracle' endpoint reference and the learned F is small, then approximation error is not a major concern; if large, it is.
from the skeptic's reading
How the review went
4 readings by deepseek-v4-flash · 13 Nov 2026
- Reader holds up
- Skeptic agrees
- Editor wrote the verdict above
- Referee major revision: 4 major objections, 9 minor
Not run on this paper: the simulated authors' defence.
Pith did not check this review's objections against the paper.
The review in full
The review in brief, machine-written
This paper identifies a genuine problem in multi-step DMD—the fake model trained on intermediate states creates a 'reference residual' that can push the student away from the teacher even when final outputs already match—and proposes a simple fix: train the fake model only on final outputs. The theoretical analysis (Proposition 1) is clean and the synthetic 2D experiments convincingly show the effect. The main experiments on four large backbones (SD3, FLUX.2, Qwen-Image) are extensive and show consistent improvements over strong baselines. The ablations confirm that the endpoint reference is the most important component. The paper is well-written and the authors are honest about the limitations (exact denoisers assumption, Gaussian analysis showing equilibrium variance can deviate).
The stress-test concern about approximation error in the fake model is valid but the paper addresses it empirically: the ablations show that removing the endpoint reference causes the largest degradation, and the endpoint reference improves even when added to TDM. The Gaussian analysis (Lemma 4) actually shows that even with an exact endpoint fake, the student's equilibrium variance can deviate from the teacher's due to the teacher target not being a pure denoiser—this is a separate issue from the reference residual, and the paper acknowledges it. The practical benefit is clear from the experiments.
The calibrated query density and precise teacher (PT) are useful engineering contributions but less central. The paper does not overclaim; it states that eDMD 'consistently outperforms' baselines, which the data support.
Who this is for: anyone working on diffusion distillation, especially DMD-based methods. The paper deserves a serious referee; the claims are well-supported and the method is likely to be adopted. I would cite it and bring it to a reading group.
Recommendation: accept with minor revisions. The main thing to check is whether the Gaussian analysis (Lemma 4) is correctly interpreted—it shows a limitation of the endpoint reference under guided/PT targets, which the paper should discuss more prominently. Also, the paper should clarify that the theoretical guarantee (Prop. 1) is for exact denoisers, and the empirical evidence shows it holds in practice.
The strongest objection
The paper's central theoretical result (Proposition 1, Sec. 3.2) proves that the reference residual R = D_q - D_ν vanishes if and only if ν = q, assuming exact conditional denoisers under a shared forward noising process. The paper acknowledges this assumption (b85, b290) and provides empirical evidence that the endpoint reference works well. However, the load-bearing step for the central claim is that training the fake model exclusively on final outputs (Eq. 7) yields F ≈ D_q, and that this approximation is sufficiently accurate to eliminate the reference residual in practice. The paper does not analyze how approximation error in the fake model (due to finite capacity, finite data, or optimization) propagates into the correction signal C = A - F. In particular, if F is a biased estimate of D_q, the residual R = D_q - F may not vanish even when ν = q, and the correction may not vanish when q = p. The paper's ablations (Tab. 3, w/o E) show that removing the endpoint reference causes the largest degradation, but this does not isolate the effect of fake model approximation error from the effect of the reference distribution mismatch. The concern is that the practical benefit of eDMD may be partly due to other components (PT, calibrated query density) or to the endpoint reference providing a better-conditioned learning signal, rather than to the elimination of the reference residual per se. The paper's own analysis of the Gaussian case (Lemma 4, App. A.5) shows that even with an exact endpoint fake, the student's equilibrium variance can deviate from the teacher's due to the teacher target A not being a pure denoiser D_p. This suggests that the theoretical guarantee is fragile under practical approximations.
The reader's weakest_assumption correctly identifies that the theoretical analysis assumes exact denoisers, and that the paper acknowledges this limitation. The reader's verdict is ACCEPT with medium correctness risk, which is appropriate. My stress-test does not identify a new load-bearing concern that would change the verdict. The concern about fake model approximation error is a natural extension of the acknowledged assumption, but the paper provides strong empirical evidence (ablations, synthetic experiments, multiple backbones) that the endpoint reference works well in practice. The concrete test I propose would further strengthen the evidence, but its absence does not invalidate the central claim. The paper is honest about limitations and provides sufficient empirical support. Therefore, the verdict should remain UNCHANGED. I agree with the reader's assessment.
The referee's report
If the results hold, eDMD provides a principled solution to a previously unrecognized objective mismatch in multi-step DMD. The theoretical analysis is clean and the empirical gains are substantial and consistent across multiple large-scale models. The method is practical (no real images needed, moderate training cost). This could become a standard component in few-step distillation pipelines.
Pith's reading of the paper (a machine's summary, not the authors')
Diffusion models generate images by gradually denoising random noise over many steps, but this is slow. Distillation trains a smaller 'student' model to do the same job in just a few steps. A popular approach, Distribution Matching Distillation (DMD), uses an extra 'fake' model to track the student's current output distribution and compares it to the original 'teacher' model to guide the student. However, when the student uses multiple denoising steps, existing methods train the fake model on the intermediate predictions from those steps, not just the final output. This paper shows that this creates a 'reference residual' that can push the student in the wrong direction, even when its final outputs already match the teacher. The proposed method, eDMD, fixes this by training the fake model only on the student's final outputs. This ensures the correction signal is only about matching the final distribution, not the intermediate ones. The paper also improves the teacher's guidance by using a short multi-step trajectory for low-noise queries and by calibrating which noise levels are used for supervision. Experiments on several large image generation models show eDMD produces better 4-step images than previous DMD methods.
Core claim. The central claim is that training the fake score model exclusively on the student's final outputs (the endpoint distribution) eliminates a 'reference residual' that distorts the distribution matching signal in multi-step DMD, and that this leads to consistently better 4-step generation quality compared to existing DMD methods. This is stated in the abstract and Sec. 4: 'eDMD trains a single shared fake model exclusively on the student's final outputs and uses it to correct every denoising step. Our theoretical analysis proves that this endpoint reference guarantees that the distribution matching term vanishes if and only if the student's final output distribution matches the teacher's, regardless of intermediate distributions.' (Abstract, b52).
The premise it rests on. The theoretical analysis (Proposition 1, Sec. 3.2) assumes 'exact conditional denoisers under a shared forward noising process' (b85, b290). In practice, the teacher and fake models are neural networks with approximation errors, and the forward process is discretized. The paper acknowledges this in the synthetic experiments (Sec. 3.3) and in the limitations of the analysis (e.g., 'assume an exact fake model, F = D_ν' in b85). The validity of the central claim in real-world settings depends on how well this idealized analysis holds when using approximate denoisers and discrete steps. The paper provides empirical evidence that it does hold, but the theoretical guarantee is conditional on this assumption.
Check this review
Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.
How often reviews like this one are right
Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.
How often objections from reviews like this one are right: not yet measured
Reviews by people
This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.
The paper
Abstract
Distribution matching distillation (DMD) has emerged as a leading paradigm for distilling diffusion models into few-step students, where an auxiliary fake score model tracks the student's distribution and the discrepancy between the teacher and fake predictions drives the student toward the teacher. In multi-step distillation, although only the student's final output is delivered, existing methods train the fake model on intermediate states or across all denoising steps. Consequently, the student's update signal becomes entangled with intermediate distributions, which can distort the matching direction and remain nonzero even after the student's final outputs already match the teacher distribution. To resolve this mismatch, we propose Endpoint Distribution Matching Distillation (eDMD), which trains a single shared fake model exclusively on the student's final outputs and uses it to correct every denoising step. Our theoretical analysis proves that this endpoint reference guarantees that the distribution matching term vanishes if and only if the student's final output distribution matches the teacher's, regardless of intermediate distributions. Beyond the fake reference, we further strengthen teacher supervision through a precise teacher (PT) that replaces single-step guided predictions with short multi-step trajectories on the final noise interval, paired with a calibrated query density for intermediate steps. Experiments on SD3-Medium, FLUX.2-klein (4B and 9B), and Qwen-Image-2.1 demonstrate that eDMD consistently outperforms existing distribution matching methods for 4-step generation.
The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.
abstract compiled · 2026-11-14
The authors' main claim, as the classifier picked it out 0.95
“Experiments on SD3-Medium, FLUX.2-klein (4B and 9B), and Qwen-Image-2.1 demonstrate that eDMD consistently outperforms existing distribution matching methods for 4-step generation.”
How the main claim is stated: asserted 1.00
What the abstract itself states, as probabilities: claims first 0.08 · states limitation 0.09 · human participants 0.02 · controlled comparison 0.73 · randomized assignment 0.01 · states quantitative result 0.35 · compares with prior results 0.95 · mentions code or data availability 0.02
The abstract, sentence by sentence
This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.
| Sentence | Role | The authors' own work | States the main conclusion | Strength |
|---|---|---|---|---|
| Distribution matching distillation (DMD) has emerged as a leading paradigm for distilling diffusion models into few-step students, where an auxiliary fake score model tracks the student's distribution and the discrepancy between the teacher and fake predictions drives the student toward the teacher. | background 1.00 | 0.04 | 0.04 | asserted 1.00 |
| In multi-step distillation, although only the student's final output is delivered, existing methods train the fake model on intermediate states or across all denoising steps. | background 0.53 | 0.06 | 0.03 | asserted 1.00 |
| Consequently, the student's update signal becomes entangled with intermediate distributions, which can distort the matching direction and remain nonzero even after the student's final outputs already match the teacher distribution. | gap 0.55 | 0.12 | 0.06 | asserted 0.67 |
| To resolve this mismatch, we propose Endpoint Distribution Matching Distillation (eDMD), which trains a single shared fake model exclusively on the student's final outputs and uses it to correct every denoising step. | contribution 0.56 | 0.93 | 0.27 | asserted 1.00 |
| Our theoretical analysis proves that this endpoint reference guarantees that the distribution matching term vanishes if and only if the student's final output distribution matches the teacher's, regardless of intermediate distributions. | contribution 0.75 | 0.94 | 0.48 | asserted 1.00 |
| Beyond the fake reference, we further strengthen teacher supervision through a precise teacher (PT) that replaces single-step guided predictions with short multi-step trajectories on the final noise interval, paired with a calibrated query density for intermediate steps. | method 0.71 | 0.89 | 0.06 | asserted 0.99 |
| Experiments on SD3-Medium, FLUX.2-klein (4B and 9B), and Qwen-Image-2.1 demonstrate that eDMD consistently outperforms existing distribution matching methods for 4-step generation. | result 1.00 | 0.96 | 0.48 | asserted 1.00 |
What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.97
What was examined, and what was not
Sources read
- arXiv listing abstract v1
Not examined
- full text
- figures and tables
- references and citing sentences
- relations to other papers
Levels run
- P abstract
Not run
- S sections
- T body sentences
- C citations
- R cross-paper relations
How the labels were made
Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-11-14: 1 request, 5,862 input tokens.
Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON
Figures (4)
Figures are the authors', reproduced from the paper.
References (68)
- [1] Donghoon Ahn, Hyoungwon Cho, Jaewon Min, Wooseok Jang, Jungwoo Kim, SeonHwa Kim, Hyun Hee Park, Ky- ong Hwan Jin, and Seungryong Kim. Self-Rectifying Diffu- sion Sampling with Perturbed-Attention Guidance. InEuro- pean Conference on Computer Vision (ECCV), 2024. 4, 6, 13 on Pith
- [2] Black Forest Labs. FLUX.2 [klein]: Towards interac- tive visual intelligence.https : / / bfl . ai / blog / flux2- klein- towards- interactive- visual- intelligence, 2026. 3, 7 on Pith
- [3] Arwen Bradley and Preetum Nakkiran. Classifier- Free Guidance is a Predictor-Corrector.arXiv preprint arXiv:2408.09000, 2024. 7 on PithPith's reviewarXiv
- [4] Huayu Chen, Kai Jiang, Kaiwen Zheng, Jianfei Chen, Hang Su, and Jun Zhu. Visual Generation Without Guidance. InIn- ternational Conference on Machine Learning (ICML), 2025. 4 on Pith
- [5] Bradley Efron. Tweedie’s formula and selection bias.Jour- nal of the American Statistical Association, 106(496):1602– 1614, 2011. 1 on Pith
- [6] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas M ¨uller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, and Robin Rombach. Scaling Rec- tified Flow Transformers for High-Resolution Image Syn... on Pith
- [7] Xiangyu Fan, Zesong Qiu, Zhuguanyu Wu, Fanzhou Wang, Zhiqian Lin, Tianxiang Ren, Dahua Lin, Ruihao Gong, and Lei Yang. Phased DMD: Few-step Distribution Match- ing Distillation via Score Matching within Subintervals. InIEEE/CVF Conference on Computer Vision and Pattern Recogni... on Pith
- [8] Xingtong Ge, Xin Zhang, Tongda Xu, Yi Zhang, Xinjie Zhang, Yan Wang, and Jun Zhang. SenseFlow: Scaling Distribution Matching for Flow-based Text-to-Image Distil- lation. InInternational Conference on Learning Representa- tions (ICLR), 2026. 4 on Pith
- [9] Jonathan Ho and Tim Salimans. Classifier-Free Diffusion Guidance.arXiv preprint arXiv:2207.12598, 2022. 2, 4, 6 9 on PithPith's reviewarXiv
- [10] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising Diffu- sion Probabilistic Models. InAdvances in Neural Informa- tion Processing Systems (NeurIPS), 2020. 2, 4 on Pith
- [11] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen- Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-Rank Adaptation of Large Language Models. InInternational Conference on Learning Representations (ICLR), 2022. 11 on Pith
- [12] Xiwei Hu, Rui Wang, Yixiao Fang, Bin Fu, Pei Cheng, and Gang Yu. ELLA: Equip diffusion models with LLM for enhanced semantic alignment.arXiv preprint arXiv:2403.05135, 2024. 7, 11 on PithPith's reviewarXiv
- [13] Yushi Huang, Xiangxin Zhou, Ruoyu Wang, Chi Zhang, Jun Zhang, and Tianyu Pang. Reinforcing few-step gener- ators via reward-tilted distribution matching.arXiv preprint arXiv:2605.26108, 2026. 4, 6, 13 on PithPith's reviewarXiv
- [14] Tero Karras, Miika Aittala, Tuomas Kynk ¨a¨anniemi, Jaakko Lehtinen, Timo Aila, and Samuli Laine. Guiding a Diffusion Model with a Bad Version of Itself. InAdvances in Neural Information Processing Systems (NeurIPS), 2024. 4, 6, 7 on Pith
- [15] Dongjun Kim, Chieh-Hsin Lai, Wei-Hsiang Liao, Naoki Mu- rata, Yuhta Takida, Toshimitsu Uesaka, Yutong He, Yuki Mitsufuji, and Stefano Ermon. Consistency Trajectory Mod- els: Learning Probability Flow ODE Trajectory of Diffu- sion. InInternational Conference on Learning Represe... on Pith
- [16] Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Matiana, Joe Penna, and Omer Levy. Pick-a-Pic: An Open Dataset of User Preferences for Text-to-Image Genera- tion. InAdvances in Neural Information Processing Systems (NeurIPS), 2023. 7 on Pith
- [17] Tuomas Kynk ¨a¨anniemi, Miika Aittala, Tero Karras, Samuli Laine, Timo Aila, and Jaakko Lehtinen. Applying Guid- ance in a Limited Interval Improves Sample and Distribution Quality in Diffusion Models. InAdvances in Neural Infor- mation Processing Systems (NeurIPS), 2024. 4 on Pith
- [18] Haoyu Li, Tingyan Wen, Lin Qi, Zhe Wu, Yihuang Chen, Xing Zhou, Lifei Zhu, Xueqian Wang, and Kai Zhang. 1.x- Distill: Breaking the Diversity, Quality, and Efficiency Bar- rier in Distribution Matching Distillation.arXiv preprint arXiv:2604.04018, 2026. 4 on PithPith's reviewarXiv
- [19] Shanchuan Lin, Anran Wang, and Xiao Yang. SDXL- Lightning: Progressive Adversarial Diffusion Distillation. arXiv preprint arXiv:2402.13929, 2024. 2, 4 on PithPith's reviewarXiv
- [20] Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maxi- milian Nickel, and Matt Le. Flow Matching for Generative Modeling. InInternational Conference on Learning Repre- sentations (ICLR), 2023. 4 on Pith
show 48 more
- [21] Dongyang Liu, Peng Gao, David Liu, Ruoyi Du, Zhen Li, Qilong Wu, Xin Jin, Sihan Cao, Shifeng Zhang, Steven Hoi, and Hongsheng Li. Decoupled DMD: CFG Augmentation as the Spear, Distribution Matching as the Shield. InIn- ternational Conference on Learning Representations (ICLR), on Pith
- [22] Haohe Liu, Zehua Chen, Yi Yuan, Xinhao Mei, Xubo Liu, Danilo Mandic, Wenwu Wang, and Mark D. Plumbley. Au- dioLDM: Text-to-Audio Generation with Latent Diffusion Models. InInternational Conference on Machine Learning (ICML), pages 21450–21474, 2023. 2 on Pith
- [23] Tao Liu, Hao Yan, Mengting Chen, Taihang Hu, Zhengrong Yue, Zihao Pan, Jinsong Lan, Xiaoyong Zhu, Ming-Ming Cheng, Bo Zheng, and Yaxing Wang. Continuous-Time Distribution Matching for Few-Step Diffusion Distillation. arXiv preprint arXiv:2605.06376, 2026. 2, 4, 5, 6, 7, 12, 13 on PithPith's reviewarXiv
- [24] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow. InInternational Conference on Learn- ing Representations (ICLR), 2023. 4 on Pith
- [25] Xingchao Liu, Xiwen Zhang, Jianzhu Ma, Jian Peng, and Qiang Liu. InstaFlow: One Step is Enough for High-Quality Diffusion-Based Text-to-Image Generation. InInternational Conference on Learning Representations (ICLR), 2024. 4 on Pith
- [26] Yijun Liu, Jie Huang, Zeyue Xue, Yuming Li, Ruizhe He, Haoran Li, Shijia Ge, and Siming Fu. HPSv3++: Scaling Re- ward Models Across the Full Spectrum of Diffusion Model Capabilities.arXiv preprint arXiv:2606.14657, 2026. 3, 7 on PitharXiv
- [27] Yuxi Liu, Haoyu Li, Yixiang Cai, Tengxu Sun, Zekun Zhang, Baole Ai, Ang Wang, Jiamang Wang, Lin Qu, Kun Yuan, and Kai Zhang. CrossDistill: Balancing Quality and Diversity via Trajectory-Level Hybrid Few-Step Distillation.arXiv preprint arXiv:2609.14725, 2026. 4 on PithPith's reviewarXiv
- [28] Ilya Loshchilov and Frank Hutter. Decoupled Weight De- cay Regularization. InInternational Conference on Learning Representations (ICLR), 2019. 11 on Pith
- [29] Cheng Lu and Yang Song. Simplifying, Stabilizing and Scal- ing Continuous-Time Consistency Models. InInternational Conference on Learning Representations (ICLR), 2025. 4 on Pith
- [30] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. DPM-Solver++: Fast Solver for Guided Sampling of Diffusion Probabilistic Models.arXiv preprint arXiv:2211.01095, 2022. 12 on PithPith's reviewarXiv
- [31] Yanzuo Lu, Yuxi Ren, Xin Xia, Shanchuan Lin, Xing Wang, Xuefeng Xiao, Andy J. Ma, Xiaohua Xie, and Jian-Huang Lai. Adversarial Distribution Matching for Diffusion Dis- tillation Towards Efficient Image and Video Synthesis. In IEEE/CVF International Conference on Computer Visio... on Pith
- [32] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent Consistency Models: Synthesizing High- Resolution Images with Few-Step Inference.arXiv preprint arXiv:2310.04378, 2023. 4 on PithPith's reviewarXiv
- [33] Weijian Luo, Tianyang Hu, Shifeng Zhang, Jiacheng Sun, Zhenguo Li, and Zhihua Zhang. Diff-Instruct: A Univer- sal Approach for Transferring Knowledge From Pre-trained Diffusion Models. InAdvances in Neural Information Pro- cessing Systems (NeurIPS), 2023. 2, 4 on Pith
- [34] Yihong Luo, Tianyang Hu, Jiacheng Sun, Yujun Cai, and Jing Tang. Learning Few-Step Diffusion Models by Tra- jectory Distribution Matching. InIEEE/CVF International Conference on Computer Vision (ICCV), 2025. 2, 4, 5, 6, 7, 10, 12 on Pith
- [35] Yuhang Ma, Xiaoshi Wu, Keqiang Sun, and Hongsheng Li. HPSv3: Towards Wide-Spectrum Human Preference Score. InIEEE/CVF International Conference on Computer Vision (ICCV), 2025. 7 10 on Pith
- [36] Chenlin Meng, Robin Rombach, Ruiqi Gao, Diederik Kingma, Stefano Ermon, Jonathan Ho, and Tim Salimans. On Distillation of Guided Diffusion Models. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 4 on Pith
- [37] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Milden- hall. DreamFusion: Text-to-3D using 2D Diffusion. InIn- ternational Conference on Learning Representations (ICLR), on Pith
- [38] Qwen Team. Qwen-Image-2.1.https://qwen.ai/ blog?id=qwen-image-2.1, 2026. 3, 7 on Pith
- [39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. InInternational Conf... on Pith
- [40] Yuxi Ren, Xin Xia, Yanzuo Lu, Jiacheng Zhang, Jie Wu, Pan Xie, Xing Wang, and Xuefeng Xiao. Hyper-SD: Trajectory Segmented Consistency Model for Efficient Image Synthe- sis. InAdvances in Neural Information Processing Systems (NeurIPS), 2024. 2, 4 on Pith
- [41] Amirmojtaba Sabour, Sanja Fidler, and Karsten Kreis. Align your flow: Scaling continuous-time flow map distillation. arXiv preprint arXiv:2506.14603, 2025. 4 on PithPith's reviewarXiv
- [42] Seyedmorteza Sadat, Otmar Hilliges, and Romann M. We- ber. Eliminating Oversaturation and Artifacts of High Guid- ance Scales in Diffusion Models. InInternational Confer- ence on Learning Representations (ICLR), 2025. 4 on Pith
- [43] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, David Fleet, and Mohammad Norouzi. Pho- torealistic Text-to-Image Diffusion Models with Deep Lan- guage... on Pith
- [44] Tim Salimans and Jonathan Ho. Progressive Distillation for Fast Sampling of Diffusion Models. InInternational Con- ference on Learning Representations (ICLR), 2022. 2, 4 on Pith
- [45] Axel Sauer, Frederic Boesel, Tim Dockhorn, Andreas Blattmann, Patrick Esser, and Robin Rombach. Fast High- Resolution Image Synthesis with Latent Adversarial Diffu- sion Distillation. InSIGGRAPH Asia, 2024. 4 on Pith
- [46] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial Diffusion Distillation. InEuropean Conference on Computer Vision (ECCV), 2024. 4 on Pith
- [47] Christoph Schuhmann. LAION-Aesthetics.https:// laion.ai/blog/laion-aesthetics/, 2022. 7 on Pith
- [48] Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Ab- hishek Kumar, Stefano Ermon, and Ben Poole. Score-Based Generative Modeling through Stochastic Differential Equa- tions. InInternational Conference on Learning Representa- tions (ICLR), 2021. 4 on Pith
- [49] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency Models. InInternational Conference on Machine Learning (ICML), 2023. 2, 4 on Pith
- [50] Team Wan, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianx- iao Yang, Jianyuan Zeng, Jiayu Wang, Jingfeng Zhang, Jin- gren Zhou, Jinkai Wang, Jixuan Chen, Kai Zhu, Kang Zhao, Keyu Yan, Lianghua Huang, Mengyang Feng, Ningyi Zhang, P... on PithPith's reviewarXiv
- [51] Ilya Tolstikhin, Sylvain Gelly, Olivier Bousquet, Carl- Johann Simon-Gabriel, and Bernhard Sch ¨olkopf. AdaGAN: Boosting Generative Models. InAdvances in Neural Infor- mation Processing Systems (NeurIPS), 2017. 3 on Pith
- [52] Fu-Yun Wang, Zhaoyang Huang, Alexander William Bergman, Dazhong Shen, Peng Gao, Michael Lingelbach, Keqiang Sun, Weikang Bian, Guanglu Song, Yu Liu, Xiao- gang Wang, and Hongsheng Li. Phased Consistency Mod- els. InAdvances in Neural Information Processing Systems (NeurIPS), 2024. 4 on Pith
- [53] Xi Wang, Nicolas Dufour, Nefeli Andreou, Marie-Paule Cani, Victoria Fern´andez Abrevaya, David Picard, and Vicky Kalogeiton. Analysis of Classifier-Free Guidance Weight Schedulers.Transactions on Machine Learning Research (TMLR), 2024. 4 on Pith
- [54] Yifei Wang, Weimin Bai, Colin Zhang, Debing Zhang, Wei- jian Luo, and He Sun. Uni-Instruct: One-step Diffusion Model through Unified Diffusion Divergence Instruction. arXiv preprint arXiv:2505.20755, 2025. 4 on PitharXiv
- [55] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. ProlificDreamer: High-Fidelity and Diverse Text-to-3D Generation with Variational Score Distillation. InAdvances in Neural Information Processing Systems (NeurIPS), 2023. 4 on Pith
- [56] Tianhe Wu, Ruibin Li, Lei Zhang, and Kede Ma. Diversity- Preserved Distribution Matching Distillation for Fast Visual Synthesis.arXiv preprint arXiv:2602.03139, 2026. 4 on PithPith's reviewarXiv
- [57] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human Pref- erence Score v2: A Solid Benchmark for Evaluating Hu- man Preferences of Text-to-Image Synthesis.arXiv preprint arXiv:2306.09341, 2023. 7 on PithPith's reviewarXiv
- [58] Zhuguanyu Wu, Ruihao Gong, Yang Yong, Yushi Huang, Xiangyu Fan, Lei Yang, Dahua Lin, and Xianglong Liu. SGMD: Score Gradient Matching Distillation for Few-Step Video Diffusion Distillation.arXiv preprint arXiv:2605.30116, 2026. 4 on PithPith's reviewarXiv
- [59] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. ImageReward: Learning and Evaluating Human Preferences for Text-to- Image Generation. InAdvances in Neural Information Pro- cessing Systems, 2023. 7 11 on Pith
- [60] Yilun Xu, Weili Nie, and Arash Vahdat. One-step Diffusion Models withf-Divergence Distribution Matching.arXiv preprint arXiv:2502.15681, 2025. 4 on PithPith's reviewarXiv
- [61] Tianwei Yin, Micha ¨el Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fr ´edo Durand, and William T. Freeman. Improved Distribution Matching Distillation for Fast Image Synthesis. InAdvances in Neural Information Processing Systems (NeurIPS), 2024. 2, 4, 5, 7 on Pith
- [62] Tianwei Yin, Micha ¨el Gharbi, Richard Zhang, Eli Shecht- man, Fr´edo Durand, William T. Freeman, and Taesung Park. One-step Diffusion with Distribution Matching Distillation. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 4, 11 on Pith
- [63] Candi Zheng and Yuan Lan. Characteristic Guidance: Non- linear Correction for Diffusion Model at Large Guidance Scale. InInternational Conference on Machine Learning (ICML), 2024. 7 on Pith
- [64] Kaiwen Zheng, Yongxin Chen, Huayu Chen, Guande He, Ming-Yu Liu, Jun Zhu, and Qinsheng Zhang. Direct Dis- criminative Optimization: Your Likelihood-Based Visual Generative Model is Secretly a GAN Discriminator. InIn- ternational Conference on Machine Learning (ICML), 2025. 4 on Pith
- [65] Mingyuan Zhou, Huangjie Zheng, Zhendong Wang, Mingzhang Yin, and Hai Huang. Score identity Distillation: Exponentially Fast Distillation of Pretrained Diffusion Mod- els for One-Step Generation. InInternational Conference on Machine Learning (ICML), 2024. 4 on Pith
- [66] Mingyuan Zhou, Yi Gu, and Zhendong Wang. Few-Step Diffusion via Score identity Distillation.arXiv preprint arXiv:2505.12674, 2025. 4 on PithPith's reviewarXiv
- [67] Mingyuan Zhou, Zhendong Wang, Huangjie Zheng, and Hai Huang. Guided Score identity Distillation for Data-Free One-Step Text-to-Image Generation. InInternational Con- ference on Learning Representations (ICLR), 2025. 4 on Pith
- [68] Zhizhuo Zhou and Shubham Tulsiani. SparseFusion: Dis- tilling View-conditioned Diffusion for 3D Reconstruction. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 4 12 Appendix A Proofs. . . . . . . . . . . . . . . . . . . . . . . . . .1 A.1 The ref... on Pith
Challenges and discussion
Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.
Come back by email
A short list of the week's newly reviewed papers. Nothing is sent until you confirm from your inbox, and every email has a link that stops it. No account.
No comments yet.
Sign in to comment