Pith. sign in

REVIEW 2 major objections 6 minor 1 cited by

Matching distributions across resolutions lets few-step cascaded diffusion run much faster without the usual quality collapse.

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 · grok-4.5

2026-07-15 13:59 UTC pith:HF7D7DOD

load-bearing objection Solid systems distillation paper that unifies logSNR-partitioned multi-res cascades with DMD-style matching; the modeling premise is imperfect but the multi-backbone evidence is strong enough to take seriously. the 2 major comments →

arxiv 2603.06136 v2 pith:HF7D7DOD submitted 2026-03-06 cs.CV

Cross-Resolution Distribution Matching for Diffusion Distillation

Feiyang Chen , Hongpeng Pan , Haonan Xu , Xinyu Duan , Zhefeng Wang , Yang Yang This is my paper
classification cs.CV
keywords diffusion distillationcross-resolutiondistribution matchinglogSNRcascaded generationfew-step samplingtext-to-imagetext-to-video
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.

Diffusion models generate high-quality images and videos, but each denoising step is expensive, and simply cutting the number of steps eventually destroys quality. Running early, high-noise steps at lower resolution would save a lot of compute, yet the same model produces different statistical distributions at different resolutions, so naive cascading looks bad. This paper introduces Cross-Resolution Distribution Matching Distillation (RMD), which partitions the denoising trajectory by log-signal-to-noise-ratio curves, maps those intervals so low- and high-resolution states sit in comparable noise regimes, and trains the student so its low-resolution outputs, after upsampling and controlled noise re-injection, match the teacher's high-resolution distribution. The result is a few-step, coarse-to-fine generator that keeps visual fidelity while delivering large speedups on both image and video backbones. A sympathetic reader cares because the method attacks the remaining efficiency wall that pure step-reduction distillation has already hit.

Core claim

The authors claim that explicitly aligning low-resolution generator distributions to the teacher's high-resolution distribution—via logSNR-based trajectory division, resolution-aware timestep mapping, reverse-KL score matching after differentiable upsampling, and mixed predicted/Gaussian noise re-injection—lets a few-step multi-resolution cascade preserve high visual fidelity while substantially accelerating sampling.

What carries the argument

Cross-Resolution Distribution Matching Distillation (RMD): logSNR-partitioned resolution intervals plus the mapping logSNR(ri) = logSNR − 2 log(ri/rK), followed by score-based reverse-KL matching of upsampled student states against the teacher and a predicted-noise re-injection rule during upsampling.

Load-bearing premise

The method assumes that the logSNR shift formula plus upsampling and mixed noise re-injection is enough to put low-resolution student states into the same denoising regime as the high-resolution teacher so that score matching truly aligns the distributions that matter.

What would settle it

Train an otherwise identical cascade that uses the same logSNR schedule and upsampling but omits the cross-resolution reverse-KL / score-matching objective; if human preference, HPS, VBench and visual quality collapse back to the naive low-res baseline while speed stays the same, the distribution-matching claim fails.

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

If this is right

  • Few-step cascaded generators can keep most of the quality of multi-step teachers while running early steps at lower resolution, cutting attention cost that grows with spatial size.
  • The same logSNR-aligned matching recipe extends from UNet image models to large Diffusion Transformers for video without architecture-specific redesign.
  • Progressive multi-stage schedules (e.g., 256→512→1024) become trainable end-to-end rather than requiring separate refiners or post-hoc cascade stages.
  • Once the distribution gap is closed, further step reduction below the usual 4–6-step floor becomes limited mainly by remaining trajectory error rather than by resolution mismatch.

Where Pith is reading between the lines

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

  • If logSNR alignment is the main lever, the same partitioning idea could be applied to other axes that change effective noise schedules (temporal length in video, latent channel count, or CFG scale).
  • The warm-up that first distills only the low-logSNR semantic stage suggests that layout errors dominate later failures; curriculum designs that lock structure early may generalize beyond this paper.
  • Predicted-noise re-injection with a tunable mix of stochasticity is a practical surrogate for full reverse-SDE simulation across resolutions and may be reusable in any cascade that must jump resolution mid-trajectory.

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

2 major / 6 minor

Summary. The paper proposes Cross-Resolution Distribution Matching Distillation (RMD), a framework that compresses a multi-step diffusion teacher into a few-step multi-resolution cascaded student. Timestep intervals are partitioned via logSNR curves; a resolution-dependent logSNR shift (Eq. 2) maps low-resolution student timesteps onto the teacher’s high-resolution schedule. Intermediate low-resolution states are projected into the teacher space by differentiable upsampling plus mixed predicted/Gaussian noise re-injection (Eqs. 4, 10), after which reverse-KL / score matching (Eqs. 6–8) aligns the projected student marginals with the teacher. A warm-up on the low-logSNR semantic stage and multi-stage schedules (K=2 or 3) are used. Experiments on SDXL, PixArt-α, SD3.5 and Wan2.1-14B report 2+2 / 3+3 NFE inference with claimed speed-ups of 33.4× and 25.6× while matching or exceeding DMD2, TDM and adversarial distillation baselines on HPS, Aes, CLIP, VBench and a small user study.

Significance. If the modeling premise holds, RMD offers a practical route past the diminishing returns of pure step distillation by exploiting the coarse-to-fine structure of the denoising trajectory. The multi-backbone image results, video extension to Wan2.1-14B, systematic ablations of RM vs. UP (Table 3), α mixing (Fig. 7), step allocation (Table 4) and multi-stage schedules (Table 5), plus a user study, constitute a solid empirical package that would be useful to practitioners seeking further acceleration of large DiT-based generators. The method is data-free (prompt-only) and re-uses standard DMD-style score matching, so it is relatively easy to re-implement on top of existing distillation codebases.

major comments (2)
  1. §3.1–3.3, Eqs. (2), (4), (10): The central claim that cascaded few-step inference preserves high-fidelity generation rests on the premise that the continuous logSNR shift of Eq. (2) (taken from SD3 scaling) together with bilinear upsampling and a fixed α-mix of predicted and Gaussian noise places low-resolution generator states into the same denoising regime as the high-resolution teacher. The paper never directly measures residual cross-resolution mismatch after this projection (e.g., score discrepancy, FID between projected student latents and teacher latents at matched logSNR, or trajectory-level statistics). End-to-end proxies (HPS, Aes, CLIP, VBench) and the RM/UP ablation in Table 3 can improve even when only the projected marginals are matched. A quantitative diagnostic of residual mismatch, or an ablation that replaces Eq. (2) with a learned or schedule-specific mapping, is neede
  2. Tables 1–2 and speed-up claims: Reported speed-ups (33.4× on SDXL, 25.6× on Wan2.1) compare NFE 2+2 / 3+3 against the multi-step teacher with CFG, while the strongest step-distillation baselines (DMD2, TDM) are evaluated at fixed high-resolution NFE=4 or 6. The paper does not report wall-clock latency of a pure high-resolution 4-step (or 6-step) DMD/TDM student versus the multi-resolution RMD student under identical hardware and batch settings, nor does it isolate the contribution of resolution reduction from the contribution of the distillation objective itself. Without that controlled comparison the claim that RMD “breaks through the sampling efficiency bottleneck of conventional step-distillation methods” remains only partially supported.
minor comments (6)
  1. §4.3 heading “Abaltion Study” is a typo; correct to “Ablation Study”.
  2. Fig. 3 caption and surrounding text refer to “Resolution-induced Shift” but do not state the exact noise schedule (variance-preserving, rectified flow, etc.) used to plot the logSNR curves; adding the schedule would improve reproducibility.
  3. Eq. (8) uses the Pseudo-Huber metric (Appendix B) while the main text writes an ℒ₂ form; a short note in the main text would avoid confusion.
  4. Table 5 reports a 3-stage schedule with logSNR thresholds −6.0/−2.5 etc., yet the default claimed in the text after Table 4 is logSNR₁=−5.0; reconcile the numbers or clarify which setting is used for the main results.
  5. Appendix C user study: number of evaluators and statistical significance of the preference rates are not reported; a brief statement would strengthen the claim.
  6. Notation: t_i := (r_i, t) is overloaded with the continuous teacher timestep t; a distinct symbol for the resolution-aware index would improve readability.

Circularity Check

0 steps flagged

No significant circularity: empirical distillation method whose objectives, schedules, and projections are standard constructions with free hyperparameters, not self-referential predictions.

full rationale

RMD is a training procedure that partitions the diffusion trajectory by logSNR curves (Eq. 1–2, citing the known SD3 resolution scaling), projects low-resolution generator states into the teacher’s high-resolution space via differentiable upsampling plus mixed predicted/Gaussian noise (Eqs. 4, 10), and minimizes a reverse-KL / score-matching objective (Eqs. 6–8) that is the standard DMD-style loss. The free parameters (logSNR cut-points, mixing coefficient α, stage weights λ_ri, warm-up) are chosen by ablation (Tables 3–5, Fig. 7) and reported; they are not fitted to a target quantity that is then re-presented as a prediction. There is no uniqueness theorem, self-citation chain that forces the central claim, or renaming of a known identity. Empirical speed/quality numbers (Tables 1–2) are measured against external baselines on public prompts and metrics; the derivation chain therefore does not reduce by construction to its own inputs. Circularity score is zero.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 3 invented entities

The central performance claim rests on standard diffusion/score-distillation machinery plus several hand-chosen schedule and mixing parameters that define when and how resolution changes. No new physical entities; the ‘invented’ pieces are algorithmic constructs whose only evidence is the paper’s own ablations and benchmarks.

free parameters (5)
  • logSNR stage thresholds (e.g. logSNR1)
    Chooses where the cascade switches resolution; primary experiments use -2.5 for K=2; ablations also try -6.0, 0.0, and multi-stage pairs. Directly controls NFE allocation and reported speedup/quality tradeoff (Tables 4–5).
  • noise re-injection mix α (with β, α²+β²=1)
    Balances predicted vs Gaussian noise in Eq. (10); ablated on PixArt-α with optimum α=0.2; training/inference α set per backbone (0.2/0.5 train, 1.0/0.9 infer).
  • resolution-aware loss weights λ_ri and λ_snr
    Balance stages in generator and fake-score objectives (Eqs. 6–9); described as resolution-aware but not given a unique first-principles value.
  • NFE split and stage count K (e.g. 2+2, 3+3, 1+1+2)
    Hand-chosen inference budgets that define the speedup claims relative to 40×2 / 50×2 teachers.
  • optimizer hyperparameters and warm-up step counts
    AdamW LRs, β2, batch size, 50/20 warm-up steps, gradient clip 1.0, Pseudo-Huber C—standard but fitted for stable distillation (Appendix B).
axioms (5)
  • domain assumption Higher resolution experiences weaker effective pixel corruption for the same noise magnitude; logSNR growth rates differ by resolution and justify partitioning the trajectory by logSNR.
    Stated in §3.1 with citations to noise-scheduling literature; underpins Eq. (2) and all stage assignments.
  • domain assumption Early high-noise steps mainly set global layout; late low-noise steps refine detail, so low-res early / high-res late is a valid compute allocation.
    Motivation in §1 citing SDEdit-style analyses; if false, cascading would destroy quality regardless of matching.
  • domain assumption Reverse KL between upsampled student marginals and teacher marginals can be optimized via score difference with a learned fake score (DMD-style).
    §3.2 imports score distillation / DMD machinery (Eqs. 5–8); validity of the approximation is assumed as in prior DMD work.
  • standard math Rectified-flow / flow-matching noise schedules and the σ_T ↔ logSNR conversion in Eq. (1) correctly describe the teacher and student processes used.
    Standard RF parameterization cited from Liu et al. / SD3; used for interval mapping.
  • ad hoc to paper Bilinear upsampling of clean latents plus controlled noise re-injection yields a differentiable map into the teacher’s high-resolution noise space suitable for distribution matching.
    Core of §3.2–3.3 (Eqs. 4, 10); justified by ablation, not derived from a uniqueness theorem.
invented entities (3)
  • Cross-Resolution Distribution Matching Distillation (RMD) objective and pipeline no independent evidence
    purpose: Train a few-step multi-resolution cascaded generator by matching teacher high-res distributions along logSNR-aligned stages.
    Named framework combining stage division, upsampling projection, fake-score training, and warm-up; evidence is internal experiments only.
  • logSNR-based cross-resolution timestep mapping (Eq. 2) no independent evidence
    purpose: Compensate resolution-induced SNR shift so student intervals correspond to teacher denoising states.
    Builds on SD3-style resolution scaling but is specialized here as the alignment device for distillation; no external validation beyond this paper’s results.
  • Predicted-noise re-injection mixture for cascade upsampling no independent evidence
    purpose: Stabilize training and inherit ODE-like trajectory while still bridging large resolution gaps.
    Introduced in §3.3 as a hybrid of pure Gaussian and pure predicted noise; α chosen by ablation.

pith-pipeline@v1.1.0-grok45 · 21673 in / 3980 out tokens · 38060 ms · 2026-07-15T13:59:03.231773+00:00 · methodology

0 comments
read the original abstract

Diffusion distillation is central to accelerating image and video generation, yet existing methods are fundamentally limited by the denoising process, where step reduction has largely saturated. Partial timestep low-resolution generation can further accelerate inference, but it suffers noticeable quality degradation due to cross-resolution distribution gaps. We propose Cross-Resolution Distribution Matching Distillation (RMD), a novel distillation framework that bridges cross-resolution distribution gaps for high-fidelity, few-step multi-resolution cascaded inference. Specifically, RMD divides the timestep intervals for each resolution using logarithmic signal-to-noise ratio (logSNR) curves, and introduces logSNR-based mapping to compensate for resolution-induced shifts. Distribution matching is conducted along resolution trajectories to reduce the gap between low-resolution generator distributions and the teacher's high-resolution distribution. In addition, a predicted-noise re-injection mechanism is incorporated during upsampling to stabilize training and improve synthesis quality. Quantitative and qualitative results show that RMD preserves high-fidelity generation while accelerating inference across various backbones. Notably, RMD achieves up to 33.4X speedup on SDXL and 25.6X on Wan2.1-14B, while preserving high visual fidelity.

Figures

Figures reproduced from arXiv: 2603.06136 by Feiyang Chen, Haonan Xu, Hongpeng Pan, Xinyu Duan, Yang Yang, Zhefeng Wang.

Figure 1
Figure 1. Figure 1: Under the same text prompt and random seed, the SDXL model produces distinct image at 512×512 and 1024×1024 resolutions, revealing a clear resolution￾dependent distribution shift. The corresponding prompts are provided in Appendix D. complexity. As a result, visual synthesis incurs substantial computational over￾head and latency, limiting the feasibility of real-time or resource-constrained applications. U… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the RMD Framework, which compresses the denoising trajectory of a pretrained diffusion model into a multi￾resolution, few-step cascaded denoising pro￾cess. 0.0 0.2 0.4 0.6 0.8 1.0 Diffusion Timestep t (0.0: Clean Image 1.0: Pure Noise) 8 6 4 2 0 2 4 6 8 logSNR Resolution-induced Shift DETAIL STAGE SEMANTIC STAGE 1024x1024 Interval 512x512 Interval Resolution: 128x128 Resolution: 256x256 Resolut… view at source ↗
Figure 4
Figure 4. Figure 4: The overall pipeline for distilling a pretrained diffusion model into a cascaded generator Gθ that performs generation across two resolution distribution spaces. \label {con:KL_divergence} \min _\theta \mathrm {KL}\Big ( p_{\theta }(\tilde {x}_{t_i}) \, p_{\theta }(x_{0} \mid \tilde {x}_{t_i}) \;\|\; p_{\varphi }(x_t) \, p_{\varphi }(x_{0} \mid x_t) \Big ), (3) where pφ and pθ denote the pretrained teacher… view at source ↗
Figure 5
Figure 5. Figure 5: Visual comparison of our model with DMD2, TDM, and the base model. All distilled models are evaluated with 6 sampling steps, while the teacher model uses 50 steps with classifier-free guidance. For fair comparison, all results are generated using identical noise seeds and text prompts. Left prompt: A basketball player soaring through the air for a thunderous slam dunk. Right prompt: A frog jumping into a b… view at source ↗
Figure 6
Figure 6. Figure 6: Visual ablation comparison of models with and without RM on SDXL at a 1024px target resolution. The corresponding prompts are provided in Appendix D. 4.3 Abaltion Study Component Ablation Analysis. Tab. 3 isolates the effects of the Cross￾Resolution Distribution Matching (RM) and Upsampling (UP) modules under a unified evaluation protocol. During training, all models are initialized at a base resolution of… view at source ↗
Figure 7
Figure 7. Figure 7: Ablation study of noise re-injection strategies. Effect of varying the mixing factor α, which controls the balance between predicted and stochastic noise. The back￾bone is Pixart-α (note that α here is part of the model name) [PITH_FULL_IMAGE:figures/full_fig_p013_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Qualitative comparison of different step allocations. Prompt: A white-haired girl in a pink sweater looks out a window in her bedroom [PITH_FULL_IMAGE:figures/full_fig_p014_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Example of the human evaluation interface used in our user study. C User Study We conducted a human evaluation to compare the generation quality of RMD and Baseline (TDM). Participants were presented with pairs of anonymous im￾ages generated by both models and asked to select the one with superior visual fidelity and text-to-image alignment. For each of the four base models (PixArt-α, SDXL, SD3.5, and Wan2… view at source ↗
Figure 10
Figure 10. Figure 10: User preference results comparing RMD and Baseline (TDM) across four base models. RMD shows a consistent advantage in human-perceived image quality and prompt alignment. architectures. The results, summarized in [PITH_FULL_IMAGE:figures/full_fig_p021_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Visual comparison on PixArt-α using identical seeds and prompts. 23 [PITH_FULL_IMAGE:figures/full_fig_p023_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Visual comparison on SD3.5 using identical seeds and prompts. 24 [PITH_FULL_IMAGE:figures/full_fig_p024_12.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Continuous-Time Distribution Matching for Few-Step Diffusion Distillation

    cs.CV 2026-05 unverdicted novelty 8.0

    CDM migrates distribution matching distillation to continuous time via dynamic random-length schedules and active off-trajectory latent alignment, yielding competitive few-step image fidelity on SD3 and Longcat-Image.

Reference graph

Works this paper leans on

53 extracted references · 10 linked inside Pith · cited by 1 Pith paper

  1. [1]

    CoRR (2025)

    Bandyopadhyay, H., Entezari, R., Scott, J., Adithyan, R., Song, Y., Jampani, V.: Sd3.5-flash: Distribution-guided distillation of generative flows. CoRR (2025)

    2025

  2. [2]

    CoRR (2025)

    Cai, X., Huang, Q., Kang, Z., Li, H., Liang, S., Ma, L., Ren, S., Wei, X., Xie, R., Zhang, T.: Longcat-video technical report. CoRR (2025)

    2025

  3. [3]

    In: ICLR (2024)

    Chen, J., Yu, J., Ge, C., Yao, L., Xie, E., Wang, Z., Kwok, J.T., Luo, P., Lu, H., Li, Z.: Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. In: ICLR (2024)

    2024

  4. [4]

    arXiv preprint arXiv:2504.07963 (2025)

    Chen, S., Ge, C., Zhang, S., Sun, P., Luo, P.: Pixelflow: Pixel-space generative models with flow. arXiv preprint arXiv:2504.07963 (2025)

    Pith/arXiv arXiv 2025

  5. [5]

    In: ICML (2023)

    Chen, T.: On the importance of noise scheduling for diffusion models. In: ICML (2023)

    2023

  6. [6]

    arXiv preprint arXiv:2403.03206 (2024)

    Esser, P., Kulal, S., Blattmann, A., Entezari, R., Müller, J., et al.: Scaling rectified flow transformers for high-resolution image synthesis. arXiv preprint arXiv:2403.03206 (2024)

    Pith/arXiv arXiv 2024

  7. [7]

    Communications of the ACM63(11), 139–144 (2020)

    Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks. Communications of the ACM63(11), 139–144 (2020)

    2020

  8. [8]

    In: NeurIPS (2020)

    Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPS (2020)

    2020

  9. [9]

    Hoogeboom, E., Heek, J., Salimans, T.: Simple diffusion: End-to-end diffusion for high resolution images

  10. [10]

    In: CVPR (2024)

    Huang, Z., He, Y., Yu, J., Zhang, F., Si, C., Jiang, Y., Zhang, Y., Wu, T., Jin, Q., Chanpaisit, N., Wang, Y., Chen, X., Wang, L., Lin, D., Qiao, Y., Liu, Z.: Vbench: Comprehensive benchmark suite for video generative models. In: CVPR (2024)

    2024

  11. [11]

    CoRRabs/2511.13649(2025)

    Jiang, D., Liu, D., Wang, Z., Wu, Q., Li, L., Li, H., Jin, X., Liu, D., Li, Z., Zhang, B., Wang, M., Hoi, S.C.H., Gao, P., Yang, H.: Distribution matching distillation meets reinforcement learning. CoRRabs/2511.13649(2025)

    Pith/arXiv arXiv 2025

  12. [12]

    In: ICLR (2024)

    Kim, D., Lai, C., Liao, W., Murata, N., Takida, Y., Uesaka, T., He, Y., Mitsu- fuji, Y., Ermon, S.: Consistency trajectory models: Learning probability flow ODE trajectory of diffusion. In: ICLR (2024)

    2024

  13. [13]

    CoRR (2024)

    Lin, S., Wang, A., Yang, X.: Sdxl-lightning: Progressive adversarial diffusion dis- tillation. CoRR (2024)

    2024

  14. [14]

    In: ICLR (2023) 15

    Lipman, Y., Chen, R.T.Q., Ben-Hamu, H., Nickel, M., Le, M.: Flow matching for generative modeling. In: ICLR (2023) 15

    2023

  15. [15]

    In: CVPR

    Liu, F., Zhang, S., Wang, X., Wei, Y., Qiu, H., Zhao, Y., Zhang, Y., Ye, Q., Wan, F.: Timestep embedding tells: It’s time to cache for video diffusion model. In: CVPR. pp. 7353–7363 (2025)

    2025

  16. [16]

    In: ICLR (2023)

    Liu, X., Huang, C., Lin, X., Wang, Y.: Flow straight and fast: Learning to generate and transfer data with rectified flow. In: ICLR (2023)

    2023

  17. [17]

    In: ICLR (2024)

    Liu, X., Zhang, X., Ma, J., Peng, J., Liu, Q.: Instaflow: One step is enough for high-quality diffusion-based text-to-image generation. In: ICLR (2024)

    2024

  18. [18]

    Lu, C., Zhou, Y., Bao, F., Chen, J., Li, C., Zhu, J.: Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. Mach. Intell. Res. (2025)

    2025

  19. [19]

    CoRRabs/2507.18569(2025)

    Lu, Y., Ren, Y., Xia, X., Lin, S., Wang, X., Xiao, X., Ma, A.J., Xie, X., Lai, J.: Adversarial distribution matching for diffusion distillation towards efficient image and video synthesis. CoRRabs/2507.18569(2025)

    Pith/arXiv arXiv 2025

  20. [20]

    In: Oh, A., Naumann, T., Globerson, A., Saenko, K., Hardt, M., Levine, S

    Luo, W., Hu, T., Zhang, S., Sun, J., Li, Z., Zhang, Z.: Diff-instruct: A universal approach for transferring knowledge from pre-trained diffusion models. In: Oh, A., Naumann, T., Globerson, A., Saenko, K., Hardt, M., Levine, S. (eds.) NIPS (2023)

    2023

  21. [21]

    In: ICLR (2025)

    Luo,Y.,Chen,X.,Qu,X.,Hu,T.,Tang,J.:Youonlysampleonce:Tamingone-step text-to-image synthesis by self-cooperative diffusion gans. In: ICLR (2025)

    2025

  22. [22]

    arXiv preprint arXiv:2503.06674 (2025)

    Luo, Y., Hu, T., Sun, J., Cai, Y., Tang, J.: Learning few-step diffusion models by trajectory distribution matching. arXiv preprint arXiv:2503.06674 (2025)

    Pith/arXiv arXiv 2025

  23. [23]

    arXiv preprint arXiv:2503.09242 (2025)

    Ma, Y., Cheng, B., Liu, S., Zhou, H., Wu, X., Wu, L., Leng, D., Yin, Y.: Nami: Efficient image generation via bridged progressive rectified flow transformers. arXiv preprint arXiv:2503.09242 (2025)

    arXiv 2025

  24. [24]

    In: ICLR (2022)

    Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J., Ermon, S.: Sdedit: Guided image synthesis and editing with stochastic differential equations. In: ICLR (2022)

    2022

  25. [25]

    In: CVPR

    Meng, C., Rombach, R., Gao, R., Kingma, D.P., Ermon, S., Ho, J., Salimans, T.: On distillation of guided diffusion models. In: CVPR. pp. 14297–14306 (2023)

    2023

  26. [26]

    In: ICCV

    Peebles, W., Xie, S.: Scalable diffusion models with transformers. In: ICCV. pp. 4172–4182 (2023)

    2023

  27. [27]

    In: ICLR (2024)

    Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: SDXL: improving latent diffusion models for high-resolution im- age synthesis. In: ICLR (2024)

    2024

  28. [28]

    In: CVPR

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR. pp. 10674–10685 (2022)

    2022

  29. [29]

    In: ICLR (2022)

    Salimans, T., Ho, J.: Progressive distillation for fast sampling of diffusion models. In: ICLR (2022)

    2022

  30. [30]

    In: ECCV (2024)

    Sauer, A., Lorenz, D., Blattmann, A., Rombach, R.: Adversarial diffusion distilla- tion. In: ECCV (2024)

    2024

  31. [31]

    Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., Schramowski, P., Kundurthy, S., Crowson, K., Schmidt, L., Kaczmarczyk, R., Jitsev, J.: Laion-5b: An open large-scale dataset for training next generation image-text models (2022)

    2022

  32. [32]

    Shen, H., Zhang, J., Xiong, B., Hu, R., Chen, S., Wan, Z., Wang, X., Zhang, Y., Gong, Z., Bao, G., Tao, C., Huang, Y., Yuan, Y., Zhang, M.: Efficient diffusion models: A survey. Trans. Mach. Learn. Res.2025(2025)

    2025

  33. [33]

    In: ICLR (2021)

    Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. In: ICLR (2021)

    2021

  34. [34]

    In: ICLR (2024)

    Song, Y., Dhariwal, P.: Improved techniques for training consistency models. In: ICLR (2024)

    2024

  35. [35]

    In: ICML (2023) 16

    Song, Y., Dhariwal, P., Chen, M., Sutskever, I.: Consistency models. In: ICML (2023) 16

    2023

  36. [36]

    In: CVPR (2025)

    Sun, K., Huang, K., Liu, X., Wu, Y., Xu, Z., Li, Z., Liu, X.: T2v-compbench: A comprehensive benchmark for compositional text-to-video generation. In: CVPR (2025)

    2025

  37. [37]

    In: NeurIPS (2023)

    Sun, K., Pan, J., Ge, Y., Li, H., Duan, H., Wu, X., Zhang, R., Zhou, A., Qin, Z., Wang, Y., Dai, J., Qiao, Y., Wang, L., Li, H.: Journeydb: A benchmark for generative image understanding. In: NeurIPS (2023)

    2023

  38. [38]

    arXiv preprint (2025)

    Team, T.H.: Hunyuanvideo: A systematic framework for large video generative models. arXiv preprint (2025)

    2025

  39. [39]

    arXiv preprint (2025)

    Team, W.: Wan: Open and advanced large-scale video generative models. arXiv preprint (2025)

    2025

  40. [40]

    Teng, J., Zheng, W., Ding, M., Hong, W., Wangni, J., Yang, Z., Tang, J.: Relay diffusion:Unifyingdiffusionprocessacrossresolutionsforimagesynthesis.In:ICLR (2024)

    2024

  41. [41]

    arXiv preprint arXiv:2503.18940 (2025)

    Tian, Y., Xia, X., Ren, Y., Lin, S., Wang, X., Xiao, X., Tong, Y., Yang, L., Cui, B.: Training-free diffusion acceleration with bottleneck sampling. arXiv preprint arXiv:2503.18940 (2025)

    Pith/arXiv arXiv 2025

  42. [42]

    CoRRabs/2303.02490(2023)

    Wang, B., Vastola, J.J.: Diffusion models generate images like painters: an analyt- ical theory of outline first, details later. CoRRabs/2303.02490(2023)

    Pith/arXiv arXiv 2023

  43. [43]

    In: NeurIPS (2024)

    Wang, F., Huang, Z., Bergman, A.W., Shen, D., Gao, P., Lingelbach, M., Sun, K., Bian, W., Song, G., Liu, Y., Wang, X., Li, H.: Phased consistency models. In: NeurIPS (2024)

    2024

  44. [44]

    CoRRabs/2602.03139(2026)

    Wu, T., Li, R., Zhang, L., Ma, K.: Diversity-preserved distribution matching dis- tillation for fast visual synthesis. CoRRabs/2602.03139(2026)

    Pith/arXiv arXiv 2026

  45. [45]

    CoRR (2023)

    Wu, X., Hao, Y., Sun, K., Chen, Y., Zhu, F., Zhao, R., Li, H.: Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. CoRR (2023)

    2023

  46. [46]

    CoRRabs/2502.15681(2025)

    Xu, Y., Nie, W., Vahdat, A.: One-step diffusion models withf-divergence distri- bution matching. CoRRabs/2502.15681(2025)

    Pith/arXiv arXiv 2025

  47. [47]

    In: NeurIPS (2024)

    Yin, T., Gharbi, M., Park, T., Zhang, R., Shechtman, E., Durand, F., Freeman, B.: Improved distribution matching distillation for fast image synthesis. In: NeurIPS (2024)

    2024

  48. [48]

    In: CVPR (2024)

    Yin, T., Gharbi, M., Zhang, R., Shechtman, E., Durand, F., Freeman, W.T., Park, T.: One-step diffusion with distribution matching distillation. In: CVPR (2024)

    2024

  49. [49]

    In: NeurIPS (2024)

    Yuan, Z., Zhang, H., Pu, L., Ning, X., Zhang, L., Zhao, T., Yan, S., Dai, G., Wang, Y.: Ditfastattn: Attention compression for diffusion transformer models. In: NeurIPS (2024)

    2024

  50. [50]

    CoRRabs/2502.05179(2025)

    Zhang, S., Li, W., Chen, S., Ge, C., Sun, P., Zhang, Y., Jiang, Y., Yuan, Z., Peng, B., Luo, P.: Flashvideo: Flowing fidelity to detail for efficient high-resolution video generation. CoRRabs/2502.05179(2025)

    arXiv 2025

  51. [51]

    CoRR abs/2508.15761(2025)

    Zhang, Y., Yang, H., Zhang, Y., Hu, Y., Zhu, F., Lin, C., Mei, X., Jiang, Y., Peng, B., Yuan, Z.: Waver: Wave your way to lifelike video generation. CoRR abs/2508.15761(2025)

    Pith/arXiv arXiv 2025

  52. [52]

    CoRR abs/2505.16864(2025)

    Zhang, Y., Xing, J., Xia, B., Liu, S., Peng, B., Tao, X., Wan, P., Lo, E., Jia, J.: Training-free efficient video generation via dynamic token carving. CoRR abs/2505.16864(2025)

    arXiv 2025

  53. [53]

    Welcome Friends

    Zhou, M., Zheng, H., Wang, Z., Yin, M., Huang, H.: Score identity distillation: Ex- ponentially fast distillation of pretrained diffusion models for one-step generation. In: ICML. vol. 235, pp. 62307–62331 (2024) 17 Appendix A Distillation and inference Algorithm Algorithm 1 presents the pseudocode of Cross-Resolution Distribution Matching Distillation (R...

    2024