Pith. sign in

REVIEW 3 major objections 4 minor 87 references

MeanFlow's one-step generation improves when training first establishes instantaneous velocity and small time-gap average velocities, then shifts to large gaps.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 20:33 UTC pith:KAU4KWYK

load-bearing objection Useful training schedule with a plausible but under-isolated mechanism; the small-gap-first weighting works, the 'large-gap corrupts v' claim is weaker than presented. the 3 major comments →

arxiv 2511.19065 v2 pith:KAU4KWYK submitted 2025-11-24 cs.CV cs.AIcs.LG

Understanding, Accelerating, and Improving MeanFlow Training

Jin-Young Kim , Hyojun Go , Lea Bogensperger , Julius Erbach , Nikolai Kalischek , Federico Tombari , Konrad Schindler , Dominik Narnhofer This is my paper
classification cs.CV cs.AIcs.LG
keywords MeanFlowinstantaneous velocityaverage velocitytraining curriculumtemporal gapflow matchingone-step generationtask affinity
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper argues that the two velocity fields learned by MeanFlow—instantaneous velocity at a single time and average velocity over a time interval—cannot be trained effectively with a fixed, uniform objective. Controlled experiments show that instantaneous velocity must be well-formed before average velocity can be learned, that large temporal gaps in average-velocity supervision destabilize instantaneous velocity, and that small-gap supervision prepares the model for later large-gap learning. The authors translate this into a curriculum: accelerate instantaneous-velocity formation with existing timestep-sampling and loss-weighting tricks, and progressively shift the average-velocity loss from short to long intervals. The result is faster convergence and better few-step generation at the same model size.

Core claim

MeanFlow's average velocity u—the key to one-step generation—is learned reliably only after the instantaneous velocity v is well formed, and the temporal gap used in u-supervision decides whether u-learning helps or hurts v. Experiments show v-pretraining accelerates u-learning while corrupting v destroys it; small-gap u-supervision builds and refines v, whereas large-gap u-supervision degrades it. Small-gap u-pretraining also yields a better launch point for large-gap u than pure v-pretraining, per task-affinity analysis. The derived curriculum—acceleration for v plus a linear shift of u-weights from small to large gaps—attains 1-NFE FID 2.87 vs 3.43 on ImageNet 256×256 with the same DiT-XL

What carries the argument

The MeanFlow identity, u(z_t, r, t) = (1/(t-r)) ∫_r^t v(z_τ, τ) dτ, which defines average velocity as the temporal integral of instantaneous velocity and underlies the training loss that supervises u against a stop-grad target computed from v; the paper's progressive weighting β(Δt, s) = 1 − s + λs(1 − Δt) that down-weights large temporal gaps early and relaxes to uniform weighting as training progresses, applied after MeanFlow's adaptive loss normalization. The identity makes the dependency of u on v explicit, and the schedule operationalizes the ordering principle.

Load-bearing premise

The curriculum's justification rests on the Task Affinity Score—cosine similarity between v-loss and u-loss gradients over 5K samples per epoch—being a reliable predictor of how smoothly large-gap u will later be learned, and on the corruption experiment's assumption that injected noise affects only the v-loss target and not the u_tgt target defined via the same velocity field.

What would settle it

Train the same DiT backbone with the proposed curriculum but reverse the schedule—supervising large-gap u from the start and adding small gaps later—with the same total compute; if the reversed schedule matches or beats the proposed order on 1-NFE FID, the central 'small-gap-first' claim is wrong. A second check: in the corruption experiment of Figure 3, measure whether the injected noise also changes the u_tgt target in Eq. 3; if u_tgt degrades directly, the observed u-learning failure could stem from a corrupted loss rather than from v being a prerequisite.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • With the same DiT-XL backbone, the proposed training reaches 1-NFE FID 2.87 on ImageNet 256×256, down from 3.43 for standard MeanFlow, and 2-NFE FID 2.64 vs 2.93.
  • The method matches the quality of the conventional MeanFlow baseline in about 2.5× fewer training epochs, or with a smaller DiT-L backbone.
  • The curriculum improves the quality of the learned instantaneous velocity itself, as shown by consistently better multi-step FID (32/64/128 NFE) when u_θ(z_t, t, t) is used as the velocity estimate.
  • Both components—velocity acceleration (MinSNR or DTD) and progressive L_u weighting—contribute independently, with the combination giving the best results across model sizes.
  • The improvements are robust to classifier-free-guidance configurations, holding for both small- and large-model CFG settings.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The ordering principle may transfer to other two-time generative models (e.g., consistency models, flow maps, shortcut models) that learn interval-integrated quantities; if so, curriculum schedules based on gap size could improve them as well.
  • The TAS proxy could be replaced by a direct measurement of optimization dynamics—e.g., how quickly large-gap u-loss decreases as a function of v quality—to make the schedule adaptive rather than fixed to iteration count.
  • The observed two-way interaction suggests that small-gap u-supervision acts as a regularizer for v; a testable extension is to see whether adding small-gap u-loss alongside v-loss helps even in settings where u is never needed for inference.
  • If the benefit is caused by v-formation speed rather than the specific schedule shape, then adaptive curricula that pause large-gap supervision until v reaches a threshold FID should match or beat the linear schedule; this is a cheap experiment to run.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper analyzes the interaction between the instantaneous velocity v and the average velocity u in MeanFlow training. Through controlled experiments it claims: (O1) v must be well-formed before u can be learned; (O2) large temporal gaps in u-supervision destabilize v, while small gaps help; (O3) gradient-affinity analysis favors pretraining with small-gap u over pure v-loss before introducing large-gap u. Based on these observations, the authors propose an accelerated MeanFlow training scheme with a progressive weighting β(Δt,s) that emphasizes small gaps early and transitions to uniform weighting, combined with standard diffusion acceleration techniques (MinSNR or DTD). The method reports improved 1-NFE FID on ImageNet 256×256 (2.87 vs. 3.43 with DiT-XL/2), faster convergence, and consistent gains across model sizes and CFG configurations. The paper includes ablations, convergence curves, and an FFHQ appendix.

Significance. If the stated causal story is correct, the paper offers both a practical training recipe and a conceptual framework for MeanFlow-style few-step generative models. The empirical gains are substantial and the ablations (Table 3, Fig. 6, FFHQ appendix) support the practical value of the proposed schedule. The paper also ships code and reproduces trends across architectures and datasets, which strengthens the engineering claim. However, the conceptual claims about learning dynamics — especially O2 and the TAS-based ordering in O3 — rely on experimental designs that do not cleanly separate the hypothesized causal mechanisms from ordinary optimization effects. Since these observations are the stated motivation for the method, the conceptual contribution needs stronger controls before the findings can be accepted as established.

major comments (3)
  1. [§4.2, Fig. 4, Appendix A] The Δt-range experiment 'Impact of u-learning on v-learning' sets the sampling ratio of t=r to zero in all conditions. Thus the large-Δt condition trains the model with only L_u and no v-loss. The observed degradation in 32-NFE FID for large Δt could result simply from the absence of any instantaneous-velocity supervision (catastrophic forgetting or drift), rather than from an active harmful interaction in which large-gap u-loss interferes with v. The actual training objective (Eqs. 5, 8, 9) always retains a v-loss component (or a small-gap u-loss that acts as a v proxy). Therefore O2 as stated — 'large-gap supervision destabilizes v' — is not established by this protocol. Please add a control that keeps the joint objective intact, for example by fixing the t=r sampling ratio constant while varying only the Δt range, and report 32-NFE FID and v-prediction error in that setting.
  2. [§4.1, Fig. 3] The corruption experiment injects Gaussian noise scaled by k·||v_t(zt|ε)|| into the target of the v-loss while training with L_MF. Because the network is shared, the injected noise corrupts the gradients for all parameters, including those that produce u. The observed degradation in u-learning may therefore be explained by noisy gradients in the shared backbone rather than by a specific causal prerequisite that well-formed v is necessary for learning u. The two-stage pretraining results (Fig. 2) support the benefit of v-pretraining, but they do not isolate the corruption mechanism. Please add a matched control in which the same amount of gradient noise is injected into the u-loss target instead of the v-loss target, or otherwise show that the effect is specific to corrupting v, e.g., by measuring v-prediction accuracy directly.
  3. [§4.3, Fig. 5, Appendix A] The Task Affinity Score (TAS) is proposed as the basis for O3, i.e., that small-Δt pretraining better prepares the model for large-Δt learning than pure v-pretraining. However, TAS is only a cosine similarity computed on 5K sampled points per epoch; the paper does not demonstrate that TAS predicts the final large-Δt performance. No final 1-NFE FID is reported for the models trained with Strategy 1 vs. Strategy 2 before the large-Δt phase. Without such a validation, the ordering prescription rests on an unvalidated proxy. The empirical ablation in Table 3 shows that progressive L_u weighting helps, but it does not isolate the 'small-Δt first versus pure v-first' comparison. Please report the actual large-Δt learning outcome for both initializations, or show a correlation between TAS and final FID across a range of schedules.
minor comments (4)
  1. [Tables 1–6] FID values are reported as single-run point estimates. Given the small differences in some comparisons (e.g., Table 3: DTD 10.96 vs. L_u weighting 10.98; Table 4: k=1 10.20 vs. k=2 11.44), the absence of error bars or multiple seeds makes it difficult to judge significance. At minimum, report the variance over 2–3 seeds for the main DiT-B results.
  2. [Fig. 3] The y-axis label 'FID (NFE=1, ε-pred)' is clear, but the curves for different k values are not annotated beyond the legend. It would help to mark the final FID values directly on the curves, as done in the FFHQ appendix, to ease reading.
  3. [§5, Eq. (8)] The notation for the adaptive loss weights (L_adp) is introduced only in Appendix B. Since the main-text equations (8) and (9) use these terms, a one-sentence pointer to the adaptive weighting mechanism (Eq. 7) in the main text would improve readability.
  4. [§6.3, Table 4] The effect of the schedule parameter k is described as 'linear is best,' but the pattern is non-monotonic (k=1 best, k=0.5 11.16, k=2 11.44, k=3 11.99). A brief discussion of why faster or slower transitions hurt — in terms of the observations in Sec. 4 — would strengthen the analysis.

Circularity Check

0 steps flagged

No significant circularity: the findings are experimental; the definitional identity and prior-work citations are not load-bearing in the claimed derivation.

full rationale

The paper's load-bearing claims (O1-O3) are supported by controlled experiments, ablations, and external benchmarks, rather than derived from definitions. The use of the identity u(zt,t,t)=v(zt,t) in Sec. 4.2 and Sec. 6 is an evaluation convention for measuring instantaneous-velocity quality, not a fitted parameter renamed as a prediction; the claim that small-gap u-loss acts as a proxy for v-learning is explicitly treated as an empirical observation and is tested across gap ranges. The DTD acceleration component is a self-citation, but it is only one of two acceleration methods tested, both evaluated in this paper's own ablations, and the progressive L_u weighting independently improves performance. The TAS analysis uses a standard gradient-cosine measure with external citations, and no uniqueness theorem is imported from the authors' prior work. The schedule parameter λ is set to normalize an expectation rather than tuned to FID, and the schedule exponent k is ablated. Potential confounds such as the Fig. 4 protocol setting the t=r sampling ratio to zero and the shared-backbone noise injection in Fig. 3 concern experimental validity, not circularity. No step reduces by construction to its own inputs.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 0 invented entities

No new physical or model entities are introduced. The paper introduces a loss-weighting schedule, which is a method rather than a postulated entity. Free parameters are limited to the normalizing constant lambda and schedule exponent k.

free parameters (2)
  • lambda (progressive weighting normalization) = 1 / E_dt[1-dt]
    Chosen to normalize the initial beta weighting to uniform expectation; depends on the dt sampling distribution.
  • k (schedule exponent) = 1
    Controls transition speed s=1-(i/T)^k; selected via ablation (Table 4) where k=1 gave best FID.
axioms (4)
  • domain assumption MeanFlow identity (Eq. 3): u = v_t - (t-r)(v_t dx u_theta + dt u_theta)
    Adopted from [18] as the training target; the paper does not re-derive its validity.
  • domain assumption Task affinity score (cosine similarity of loss gradients) is a valid predictor of learning dynamics
    Used in Sec 4.3 to decide training order; not independently validated.
  • standard math Small-gap u-loss approximates v-loss because u(z,t,t)=v(z,t)
    Direct consequence of Eq. 2; used in Sec 4.2 to interpret experiments.
  • domain assumption Observations from DiT-B/4 on ImageNet and FFHQ generalize to DiT-XL/L/M
    The main scaling experiments use these larger architectures; no proof of transfer.

pith-pipeline@v1.3.0-alltime-deepseek · 18547 in / 13225 out tokens · 130044 ms · 2026-08-03T20:33:35.012019+00:00 · methodology

0 comments
read the original abstract

MeanFlow promises high-quality generative modeling in few steps, by jointly learning instantaneous and average velocity fields. Yet, the underlying training dynamics remain unclear. We analyze the interaction between the two velocities and find: (i) well-established instantaneous velocity is a prerequisite for learning average velocity; (ii) learning of instantaneous velocity benefits from average velocity when the temporal gap is small, but degrades as the gap increases; and (iii) task-affinity analysis indicates that smooth learning of large-gap average velocities, essential for one-step generation, depends on the prior formation of accurate instantaneous and small-gap average velocities. Guided by these observations, we design an effective training scheme that accelerates the formation of instantaneous velocity, then shifts emphasis from short- to long-interval average velocity. Our enhanced MeanFlow training yields faster convergence and significantly better few-step generation: With the same DiT-XL backbone, our method reaches an impressive FID of 2.87 on 1-NFE ImageNet 256x256, compared to 3.43 for the conventional MeanFlow baseline. Alternatively, our method matches the performance of the MeanFlow baseline with 2.5x shorter training time, or with a smaller DiT-L backbone.

Figures

Figures reproduced from arXiv: 2511.19065 by Dominik Narnhofer, Federico Tombari, Hyojun Go, Jin-Young Kim, Julius Erbach, Konrad Schindler, Lea Bogensperger, Nikolai Kalischek.

Figure 1
Figure 1. Figure 1: Our novel, enhanced training strategy reaches the per [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 4
Figure 4. Figure 4: Impact of ∆t of u-learning on v-learning. 32-NFE FID after 40 epochs of u finetuning across different ∆t ranges, starting from either random initialization (blue) or v-pretrained model (orange, 40 epochs). Small ∆t enables constructing and improving v, while large ∆t degrades pretrained v. The green line denotes the performance of the v-pretrained model. 0.1-0.3 0.3-0.5 0.5-0.7 0.7-0.9 t = t r 0.3 0.4 0.5 … view at source ↗
Figure 5
Figure 5. Figure 5: Task affinity between v- and u-learning across ∆t ranges. Small-∆t u-pretraining achieves higher affinity for large ∆t compared to v-pretraining, providing a better regime for learn￾ing large-gap average velocity with instantaneous velocity. words, a corrupted instantaneous velocity makes learning of the average velocity a lot harder. Implication. The two experiments above reveal symmet￾ric dependencies: u… view at source ↗
Figure 6
Figure 6. Figure 6: Convergence speed comparison between MeanFlow and our methods across model sizes. 30 epochs 60 epochs 120 epochs 240 epochs 30 epochs 60 epochs 120 epochs 240 epochs MeanFlow-XL MeanFlow-XL + Ours-DTD [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative comparison of generated samples across training epochs on DiT-XL/2. only modify the sampling distribution and leave the loss weighting scheme intact, thus preserving compatibility with MeanFlow’s adaptive design. Therefore, due to its consis￾tent performance across model sizes, we select DTD as our primary acceleration method. ImageNet 256×256 benchmark. We scale up our method with DTD to the D… view at source ↗
Figure 9
Figure 9. Figure 9: Corruption in v-learning disrupts u-learning. 1- NFE FID when training with LMF while injecting Gaussian noise scaled by k·∥vt(zt|ϵ)∥ into the target velocity of Lv. Even small noise (k = 0.03) disrupts v-learning and severely degrades u￾learning performance compared to clean training (k = 0). 0.1-0.3 0.3-0.5 0.5-0.7 0.7-0.9 t = t r 10 20 50 100 200 FID (NFE=32, v-pred) v-150: 24.27 v-0 v-150 [PITH_FULL_I… view at source ↗
Figure 10
Figure 10. Figure 10: Impact of ∆t of u-learning on v-learning. 32-NFE FID after 150 epochs of u finetuning across different ∆t ranges, starting from either random initialization (blue) or v-pretrained model (orange, 150 epochs). Small ∆t enables constructing and improving v, while large ∆t degrades pretrained v. The green line denotes the performance of the v-pretrained model. Impact of u-Learning on v-Learning [PITH_FULL_IM… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

87 extracted references · 12 linked inside Pith

  1. [1]

    Krea Realtime 14B: Real-time video generation,

    Krea AI. Krea Realtime 14B: Real-time video generation,

  2. [2]

    Building nor- malizing flows with stochastic interpolants

    Michael S Albergo and Eric Vanden-Eijnden. Building nor- malizing flows with stochastic interpolants. InInternational Conference on Learning Representations (ICLR), 2023. 1, 2, 3

    2023

  3. [3]

    All are worth words: A ViT backbone for diffusion models

    Fan Bao, Shen Nie, Kaiwen Xue, Yue Cao, Chongxuan Li, Hang Su, and Jun Zhu. All are worth words: A ViT backbone for diffusion models. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 6

    2023

  4. [4]

    FLUX.1 Kontext: Flow matching for in-context image generation and editing in latent space.preprint arXiv:2506.15742, 2025

    Stephen Batifol, Andreas Blattmann, Frederic Boesel, Sak- sham Consul, Cyril Diagne, Tim Dockhorn, Jack English, Zion English, Patrick Esser, Sumith Kulal, et al. FLUX.1 Kontext: Flow matching for in-context image generation and editing in latent space.preprint arXiv:2506.15742, 2025. 1

    Pith/arXiv arXiv 2025

  5. [5]

    TRACT: Denoising diffusion models with tran- sitive closure time-distillation.preprint arXiv:2303.04248,

    David Berthelot, Arnaud Autef, Jierui Lin, Dian Ang Yap, Shuangfei Zhai, Siyuan Hu, Daniel Zheng, Walter Talbott, and Eric Gu. TRACT: Denoising diffusion models with tran- sitive closure time-distillation.preprint arXiv:2303.04248,

    Pith/arXiv arXiv

  6. [6]

    Flow map matching.preprint arXiv:2406.07507,

    Nicholas M Boffi, Michael S Albergo, and Eric Vanden- Eijnden. Flow map matching.preprint arXiv:2406.07507,

    Pith/arXiv arXiv

  7. [7]

    How to build a consistency model: Learning flow maps via self-distillation.Advances in Neural Information Processing Systems (NeurIPS), 2025

    Nicholas M Boffi, Michael S Albergo, and Eric Vanden- Eijnden. How to build a consistency model: Learning flow maps via self-distillation.Advances in Neural Information Processing Systems (NeurIPS), 2025. 1, 3

    2025

  8. [8]

    Large scale GAN training for high fidelity natural image synthe- sis

    Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale GAN training for high fidelity natural image synthe- sis. InInternational Conference on Learning Representa- tions (ICLR), 2019. 6

    2019

  9. [9]

    MaskGIT: Masked generative image transformer

    Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. MaskGIT: Masked generative image transformer. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 6

    2022

  10. [10]

    Perception priori- tized training of diffusion models

    Jooyoung Choi, Jungbeom Lee, Chaehun Shin, Sungwon Kim, Hyunwoo Kim, and Sungroh Yoon. Perception priori- tized training of diffusion models. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 5

    2022

  11. [11]

    Imagenet: A large-scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2009. 2, 3, 6

    2009

  12. [12]

    Diffusion models beat gans on image synthesis.Advances in Neural Informa- tion Processing Systems (NeurIPS), 2021

    Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis.Advances in Neural Informa- tion Processing Systems (NeurIPS), 2021. 6

    2021

  13. [13]

    GENIE: Higher-order denoising diffusion solvers.Advances in Neu- ral Information Processing Systems (NeurIPS), 2022

    Tim Dockhorn, Arash Vahdat, and Karsten Kreis. GENIE: Higher-order denoising diffusion solvers.Advances in Neu- ral Information Processing Systems (NeurIPS), 2022. 1

    2022

  14. [14]

    Tam- ing transformers for high-resolution image synthesis

    Patrick Esser, Robin Rombach, and Bjorn Ommer. Tam- ing transformers for high-resolution image synthesis. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 6

    2021

  15. [15]

    Efficiently identifying task groupings for multi-task learning.Advances in Neural Information Pro- cessing Systems (NeurIPS), 2021

    Chris Fifty, Ehsan Amid, Zhe Zhao, Tianhe Yu, Rohan Anil, and Chelsea Finn. Efficiently identifying task groupings for multi-task learning.Advances in Neural Information Pro- cessing Systems (NeurIPS), 2021. 5

    2021

  16. [16]

    One step diffusion via shortcut models

    Kevin Frans, Danijar Hafner, Sergey Levine, and Pieter Abbeel. One step diffusion via shortcut models. InInter- national Conference on Learning Representations (ICLR),

  17. [17]

    Wan-S2V: Audio-driven cinematic video gener- ation.preprint arXiv:2508.18621, 2025

    Xin Gao, Li Hu, Siqi Hu, Mingyang Huang, Chaonan Ji, Dechao Meng, Jinwei Qi, Penchong Qiao, Zhen Shen, Yafei Song, et al. Wan-S2V: Audio-driven cinematic video gener- ation.preprint arXiv:2508.18621, 2025. 1

    Pith/arXiv arXiv 2025

  18. [18]

    Mean flows for one-step genera- tive modeling.Advances in Neural Information Processing Systems (NeurIPS), 2025

    Zhengyang Geng, Mingyang Deng, Xingjian Bai, J Zico Kolter, and Kaiming He. Mean flows for one-step genera- tive modeling.Advances in Neural Information Processing Systems (NeurIPS), 2025. 2, 3, 6, 8, 1

    2025

  19. [19]

    Consistency models made easy.Inter- national Conference on Learning Representations (ICLR),

    Zhengyang Geng, Ashwini Pokle, William Luo, Justin Lin, and J Zico Kolter. Consistency models made easy.Inter- national Conference on Learning Representations (ICLR),

  20. [20]

    Addressing nega- tive transfer in diffusion models.Advances in Neural Infor- mation Processing Systems (NeurIPS), 2023

    Hyojun Go, Yunsung Lee, Seunghyun Lee, Shinhyeok Oh, Hyeongdon Moon, and Seungtaek Choi. Addressing nega- tive transfer in diffusion models.Advances in Neural Infor- mation Processing Systems (NeurIPS), 2023. 2, 5, 6

    2023

  21. [21]

    VIST3A: Text-to-3d by stitching a multi-view reconstruction network to a video generator.preprint arXiv:2510.13454, 2025

    Hyojun Go, Dominik Narnhofer, Goutam Bhat, Prune Truong, Federico Tombari, and Konrad Schindler. VIST3A: Text-to-3d by stitching a multi-view reconstruction network to a video generator.preprint arXiv:2510.13454, 2025. 1

    arXiv 2025

  22. [22]

    SplatFlow: Multi- view rectified flow model for 3d gaussian splatting synthesis

    Hyojun Go, Byeongjun Park, Jiho Jang, Jin-Young Kim, Soonwoo Kwon, and Changick Kim. SplatFlow: Multi- view rectified flow model for 3d gaussian splatting synthesis. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 1

    2025

  23. [23]

    BOOT: Data-free distillation of denois- ing diffusion models with bootstrapping

    Jiatao Gu, Shuangfei Zhai, Yizhe Zhang, Lingjie Liu, and Joshua M Susskind. BOOT: Data-free distillation of denois- ing diffusion models with bootstrapping. InICML Workshop on Structured Probabilistic Inference & Generative Model- ing, 2023. 2

    2023

  24. [24]

    STARFlow: Scaling latent normalizing flows for high- resolution image synthesis.Advances in Neural Information Processing Systems (NeurIPS), 2025

    Jiatao Gu, Tianrong Chen, David Berthelot, Huangjie Zheng, Yuyang Wang, Ruixiang Zhang, Laurent Dinh, Miguel Angel Bautista, Josh Susskind, and Shuangfei Zhai. STARFlow: Scaling latent normalizing flows for high- resolution image synthesis.Advances in Neural Information Processing Systems (NeurIPS), 2025. 6

    2025

  25. [25]

    Efficient diffu- sion training via min-SNR weighting strategy

    Tiankai Hang, Shuyang Gu, Chen Li, Jianmin Bao, Dong Chen, Han Hu, Xin Geng, and Baining Guo. Efficient diffu- sion training via min-SNR weighting strategy. InIEEE/CVF International Conference on Computer Vision (ICCV), 2023. 1, 2, 5, 6

    2023

  26. [26]

    Mul- tistep consistency models.preprint arXiv:2403.06807, 2024

    Jonathan Heek, Emiel Hoogeboom, and Tim Salimans. Mul- tistep consistency models.preprint arXiv:2403.06807, 2024. 1, 3

    Pith/arXiv arXiv 2024

  27. [27]

    GANs trained by a two time-scale update rule converge to a local Nash equi- librium.Advances in Neural Information Processing Systems (NeurIPS), 2017

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local Nash equi- librium.Advances in Neural Information Processing Systems (NeurIPS), 2017. 6

    2017

  28. [28]

    Denoising diffu- sion probabilistic models.Advances in Neural Information Processing Systems (NeurIPS), 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffu- sion probabilistic models.Advances in Neural Information Processing Systems (NeurIPS), 2020. 1, 2 9

    2020

  29. [29]

    Sim- ple diffusion: End-to-end diffusion for high resolution im- ages

    Emiel Hoogeboom, Jonathan Heek, and Tim Salimans. Sim- ple diffusion: End-to-end diffusion for high resolution im- ages. InInternational Conference on Machine Learning (ICML), 2023. 6

    2023

  30. [30]

    CMT: Mid-training for efficient learning of consistency, mean flow, and flow map models.preprint arXiv:2509.24526, 2025

    Zheyuan Hu, Chieh-Hsin Lai, Yuki Mitsufuji, and Ste- fano Ermon. CMT: Mid-training for efficient learning of consistency, mean flow, and flow map models.preprint arXiv:2509.24526, 2025. 3

    arXiv 2025

  31. [31]

    Scaling up GANs for text-to-image synthesis

    Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scaling up GANs for text-to-image synthesis. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 6

    2023

  32. [32]

    Elucidating the design space of diffusion-based generative models.Advances in Neural Information Processing Systems (NeurIPS), 2022

    Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models.Advances in Neural Information Processing Systems (NeurIPS), 2022. 1, 2

    2022

  33. [33]

    Analyzing and improving the training dynamics of diffusion models

    Tero Karras, Miika Aittala, Jaakko Lehtinen, Janne Hellsten, Timo Aila, and Samuli Laine. Analyzing and improving the training dynamics of diffusion models. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 5

    2024

  34. [34]

    Consistency trajectory mod- els: Learning probability flow ode trajectory of diffusion

    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 diffusion. InInternational Conference on Learning Representations (ICLR), 2024. 1, 3

    2024

  35. [35]

    Denoising task difficulty-based curriculum for training diffusion models

    Jin-Young Kim, Hyojun Go, Soonwoo Kwon, and Hyun- Gyoon Kim. Denoising task difficulty-based curriculum for training diffusion models. InInternational Conference on Learning Representations (ICLR), 2025. 2, 6, 1

    2025

  36. [36]

    ACT-Diffusion: Efficient adversarial consistency training for one-step diffusion models

    Fei Kong, Jinhao Duan, Lichao Sun, Hao Cheng, Renjing Xu, Hengtao Shen, Xiaofeng Zhu, Xiaoshuang Shi, and Kaidi Xu. ACT-Diffusion: Efficient adversarial consistency training for one-step diffusion models. InIEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR),

  37. [37]

    Autoregressive image generation without vec- tor quantization.Advances in Neural Information Processing Systems (NeurIPS), 2024

    Tianhong Li, Yonglong Tian, He Li, Mingyang Deng, and Kaiming He. Autoregressive image generation without vec- tor quantization.Advances in Neural Information Processing Systems (NeurIPS), 2024. 6

    2024

  38. [38]

    Connecting consistency distillation to score distillation for text-to-3d generation

    Zongrui Li, Minghui Hu, Qian Zheng, and Xudong Jiang. Connecting consistency distillation to score distillation for text-to-3d generation. InEuropean Conference on Computer Vision (ECCV), 2024. 2

    2024

  39. [39]

    Flow matching for generative mod- eling

    Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximil- ian Nickel, and Matt Le. Flow matching for generative mod- eling. InInternational Conference on Learning Representa- tions (ICLR), 2023. 1, 2, 3

    2023

  40. [40]

    Flow straight and fast: Learning to generate and transfer data with rectified flow

    Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. InInternational Conference on Learning Rep- resentations (ICLR), 2023. 1, 2, 3

    2023

  41. [41]

    See further when clear: Curriculum consistency model

    Yunpeng Liu, Boxiao Liu, Yi Zhang, Xingzhong Hou, Guan- glu Song, Yu Liu, and Haihang You. See further when clear: Curriculum consistency model. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 2

    2025

  42. [42]

    Simplifying, stabilizing and scal- ing continuous-time consistency models.International Con- ference on Learning Representations (ICLR), 2025

    Cheng Lu and Yang Song. Simplifying, stabilizing and scal- ing continuous-time consistency models.International Con- ference on Learning Representations (ICLR), 2025. 1, 3

    2025

  43. [43]

    DPM-Solver: A fast ODE solver for dif- fusion probabilistic model sampling in around 10 steps.Ad- vances in Neural Information Processing Systems (NeurIPS),

    Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. DPM-Solver: A fast ODE solver for dif- fusion probabilistic model sampling in around 10 steps.Ad- vances in Neural Information Processing Systems (NeurIPS),

  44. [44]

    PM-Solver++: Fast solver for guided sam- pling of diffusion probabilistic models.Machine Intelligence Research, 22:730–751, 2025

    Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. PM-Solver++: Fast solver for guided sam- pling of diffusion probabilistic models.Machine Intelligence Research, 22:730–751, 2025. 1

    2025

  45. [45]

    Knowledge distillation in iterative generative models for improved sampling speed

    Eric Luhman and Troy Luhman. Knowledge distillation in iterative generative models for improved sampling speed. preprint arXiv:2101.02388, 2021. 1, 2

    Pith/arXiv arXiv 2021

  46. [46]

    Latent consistency models: Synthesizing high-resolution images with few-step inference.preprint arXiv:2310.04378, 2023

    Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference.preprint arXiv:2310.04378, 2023. 3

    Pith/arXiv arXiv 2023

  47. [47]

    Diff-Instruct: A universal approach for transferring knowledge from pre-trained diffu- sion models.Advances in Neural Information Processing Systems (NeurIPS), 2023

    Weijian Luo, Tianyang Hu, Shifeng Zhang, Jiacheng Sun, Zhenguo Li, and Zhihua Zhang. Diff-Instruct: A universal approach for transferring knowledge from pre-trained diffu- sion models.Advances in Neural Information Processing Systems (NeurIPS), 2023. 1, 2

    2023

  48. [48]

    SiT: Explor- ing flow and diffusion-based generative models with scalable interpolant transformers

    Nanye Ma, Mark Goldstein, Michael S Albergo, Nicholas M Boffi, Eric Vanden-Eijnden, and Saining Xie. SiT: Explor- ing flow and diffusion-based generative models with scalable interpolant transformers. InEuropean Conference on Com- puter Vision (ECCV), 2024. 1, 6

    2024

  49. [49]

    On distillation of guided diffusion models

    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. 1, 2

    2023

  50. [50]

    Denoising task routing for diffusion models

    Byeongjun Park, Sangmin Woo, Hyojun Go, Jin-Young Kim, and Changick Kim. Denoising task routing for diffusion models. InInternational Conference on Learning Represen- tations (ICLR), 2024. 6

    2024

  51. [51]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers. InIEEE/CVF International Conference on Computer Vision (ICCV), 2023. 3, 6

    2023

  52. [52]

    Hierarchical text-conditional image gener- ation with clip latents.preprint arXiv:2204.06125, 2022

    Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image gener- ation with clip latents.preprint arXiv:2204.06125, 2022. 2

    Pith/arXiv arXiv 2022

  53. [53]

    High-resolution image syn- thesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj¨orn Ommer. High-resolution image syn- thesis with latent diffusion models. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 6

    2022

  54. [54]

    Align your steps: Optimizing sampling schedules in diffusion models

    Amirmojtaba Sabour, Sanja Fidler, and Karsten Kreis. Align your steps: Optimizing sampling schedules in diffusion models. InInternational Conference on Machine Learning (ICML), 2024. 1

    2024

  55. [55]

    Align your flow: Scaling continuous-time flow map distil- lation.Advances in Neural Information Processing Systems (NeurIPS), 2025

    Amirmojtaba Sabour, Sanja Fidler, and Karsten Kreis. Align your flow: Scaling continuous-time flow map distil- lation.Advances in Neural Information Processing Systems (NeurIPS), 2025. 1, 3 10

    2025

  56. [56]

    Progressive distillation for fast sampling of diffusion models

    Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. InInternational Confer- ence on Learning Representations (ICLR), 2022. 1, 2

    2022

  57. [57]

    Multistep distillation of diffusion models via moment matching.Advances in Neural Information Process- ing Systems (NeurIPS), 2024

    Tim Salimans, Thomas Mensink, Jonathan Heek, and Emiel Hoogeboom. Multistep distillation of diffusion models via moment matching.Advances in Neural Information Process- ing Systems (NeurIPS), 2024. 2

    2024

  58. [58]

    StyleGAN- XL: Scaling stylegan to large diverse datasets

    Axel Sauer, Katja Schwarz, and Andreas Geiger. StyleGAN- XL: Scaling stylegan to large diverse datasets. InACM Sig- Graph, 2022. 6

    2022

  59. [59]

    Fast high- resolution image synthesis with latent adversarial diffusion distillation

    Axel Sauer, Frederic Boesel, Tim Dockhorn, Andreas Blattmann, Patrick Esser, and Robin Rombach. Fast high- resolution image synthesis with latent adversarial diffusion distillation. InACM SigGraph Asia, 2024. 1, 2

    2024

  60. [60]

    Adversarial diffusion distillation

    Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. InEuropean Conference on Computer Vision (ECCV), 2024. 1, 2

    2024

  61. [61]

    Deep unsupervised learning using nonequilibrium thermodynamics

    Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. InInternational Confer- ence on Machine Learning (ICML), 2015. 1, 2

    2015

  62. [62]

    Denois- ing diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denois- ing diffusion implicit models. InInternational Conference on Learning Representations (ICLR), 2021. 1

    2021

  63. [63]

    Improved techniques for training consistency models.International Conference on Learning Representations (ICLR), 2024

    Yang Song and Prafulla Dhariwal. Improved techniques for training consistency models.International Conference on Learning Representations (ICLR), 2024. 1, 3, 6

    2024

  64. [64]

    Generative modeling by esti- mating gradients of the data distribution.Advances in Neural Information Processing Systems (NeurIPS), 2019

    Yang Song and Stefano Ermon. Generative modeling by esti- mating gradients of the data distribution.Advances in Neural Information Processing Systems (NeurIPS), 2019. 1, 2

    2019

  65. [65]

    Score-based generative modeling through stochastic differential equa- tions

    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 Represen- tations (ICLR), 2021. 2

    2021

  66. [66]

    Consistency models

    Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. InInternational Conference on Machine Learning (ICML), 2023. 1, 2

    2023

  67. [67]

    Which tasks should be learned together in multi-task learning? InInternational Conference on Machine Learning (ICML), 2020

    Trevor Standley, Amir Zamir, Dawn Chen, Leonidas Guibas, Jitendra Malik, and Silvio Savarese. Which tasks should be learned together in multi-task learning? InInternational Conference on Machine Learning (ICML), 2020. 5

    2020

  68. [68]

    Physics informed distillation for diffusion models.Transac- tions on Machine Learning Research (TMLR), 2024

    Joshua Tian Jin Tee, Kang Zhang, Hee Suk Yoon, Dhanan- jaya Nagaraja Gowda, Chanwoo Kim, and Chang D Yoo. Physics informed distillation for diffusion models.Transac- tions on Machine Learning Research (TMLR), 2024. 2

    2024

  69. [69]

    Visual autoregressive modeling: Scalable image generation via next-scale prediction.Advances in Neural In- formation Processing Systems (NeurIPS), 2024

    Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Li- wei Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction.Advances in Neural In- formation Processing Systems (NeurIPS), 2024. 6

    2024

  70. [70]

    Wan: Open and advanced large-scale video generative models.preprint arXiv:2503.20314, 2025

    Team Wan, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianx- iao Yang, et al. Wan: Open and advanced large-scale video generative models.preprint arXiv:2503.20314, 2025. 1

    Pith/arXiv arXiv 2025

  71. [71]

    Sta- ble consistency tuning: Understanding and improving con- sistency models.preprint arXiv:2410.18958, 2024

    Fu-Yun Wang, Zhengyang Geng, and Hongsheng Li. Sta- ble consistency tuning: Understanding and improving con- sistency models.preprint arXiv:2410.18958, 2024. 1, 3

    Pith/arXiv arXiv 2024

  72. [72]

    Phased consistency models.Advances in Neural Information Pro- cessing Systems (NeurIPS), 2024

    Fu-Yun Wang, Zhaoyang Huang, Alexander Bergman, Dazhong Shen, Peng Gao, Michael Lingelbach, Keqiang Sun, Weikang Bian, Guanglu Song, Yu Liu, et al. Phased consistency models.Advances in Neural Information Pro- cessing Systems (NeurIPS), 2024. 1, 3

    2024

  73. [73]

    A closer look at time steps is worthy of triple speed-up for diffusion model training

    Kai Wang, Mingjia Shi, Yukun Zhou, Zekai Li, Zhihang Yuan, Yuzhang Shang, Xiaojiang Peng, Hanwang Zhang, and Yang You. A closer look at time steps is worthy of triple speed-up for diffusion model training. InIEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR),

  74. [74]

    Transition mod- els: Rethinking the generative learning objective.preprint arXiv:2509.04394, 2025

    Zidong Wang, Yiyuan Zhang, Xiaoyu Yue, Xiangyu Yue, Yangguang Li, Wanli Ouyang, and Lei Bai. Transition mod- els: Rethinking the generative learning objective.preprint arXiv:2509.04394, 2025. 1, 3

    Pith/arXiv arXiv 2025

  75. [75]

    Qwen-image technical report.preprint arXiv:2508.02324, 2025

    Chenfei Wu, Jiahao Li, Jingren Zhou, Junyang Lin, Kaiyuan Gao, Kun Yan, Sheng-ming Yin, Shuai Bai, Xiao Xu, Yilei Chen, et al. Qwen-image technical report.preprint arXiv:2508.02324, 2025. 1

    Pith/arXiv arXiv 2025

  76. [76]

    Im- proved distribution matching distillation for fast image syn- thesis.Advances in Neural Information Processing Systems (NeurIPS), 2024

    Tianwei Yin, Micha ¨el Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and Bill Freeman. Im- proved distribution matching distillation for fast image syn- thesis.Advances in Neural Information Processing Systems (NeurIPS), 2024. 1, 2

    2024

  77. [77]

    One-step diffusion with distribution matching distillation

    Tianwei Yin, Micha ¨el Gharbi, Richard Zhang, Eli Shecht- man, Fredo 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. 1, 2

    2024

  78. [78]

    Improving consistency distilla- tion with rectified trajectories

    Han Zhang and Fan Cheng. Improving consistency distilla- tion with rectified trajectories. InInternational Conference on Artificial Neural Networks (ICANN), 2025. 2

    2025

  79. [79]

    AlphaFlow: Understanding and improving MeanFlow models.preprint arXiv:2510.20771, 2025

    Huijie Zhang, Aliaksandr Siarohin, Willi Menapace, Michael Vasilkovsky, Sergey Tulyakov, Qing Qu, and Ivan Skorokhodov. AlphaFlow: Understanding and improving MeanFlow models.preprint arXiv:2510.20771, 2025. 3

    arXiv 2025

  80. [80]

    Fast sampling of dif- fusion models with exponential integrator

    Qinsheng Zhang and Yongxin Chen. Fast sampling of dif- fusion models with exponential integrator. InInternational Conference on Learning Representations (ICLR), 2023. 1

    2023

Showing first 80 references.