Pith. sign in

REVIEW 4 major objections 4 minor 2 cited by

CAT Pruning: Cluster-Aware Token Pruning For Text-to-Image Diffusion Models

T0 review · 4 major / 4 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read Token pruning speeds text-to-image diffusion up to 2.15x

desk verdict Real 2x speedups from a new token-selection recipe, but the 'comparable quality' claim is overstated and Proposition 1 doesn't prove what it says. read the letter →

arxiv 2502.00433 v1 pith:R4RBLQFE submitted 2025-02-01 cs.CV

Xinle Cheng , Zhuoming Chen , Zhihao Jia This is my paper · ORCID
classification cs.CV
keywords tokenpruningdiffusionmodelstext-to-imagecache-and-reuserelativenoisemagnitudespatialclusteringinferenceacceleration
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

CAT Pruning accelerates text-to-image diffusion models by updating only a third of image tokens at each denoising step and reusing cached hidden states for the remaining two-thirds. The paper claims this reduces multiply-accumulate operations by roughly 50 percent at 28 sampling steps and 60 percent at 50 steps, giving end-to-end speedups up to 2.15x on Stable Diffusion v3 and Pixart-Σ while keeping CLIP Score within about 0.3 points of the full model. The method ranks tokens by relative noise magnitude, adds KMeans clustering with positional encoding to preserve spatial coherence, and uses an exponentially weighted frequency counter to keep token selection balanced over time. A sympathetic reader would care because it attacks the per-kernel compute cost of diffusion transformers directly, without retraining or changing the sampling schedule.

What carries the argument

The load-bearing identity is the relative noise magnitude $n_t - n_{t_0}$, used as a per-token importance score: tokens whose predicted noise has moved most since a reference step are the ones updated, while the rest reuse their previous hidden state. Selection is made spatially aware by adding a normalized positional encoding to the noise vector and running KMeans once at step $t_0+1$; a cheap graph pooling layer scores clusters, and a top-$k$ selection picks clusters and then tokens within them. An exponentially weighted moving average of each token's selection frequency implements the exploration-exploitation tradeoff, preventing the same tokens from being starved or overused. The reuse of hidden states in noise space is treated as a special case of interpolating output noises across two consecutive steps, which connects it to high-order ODE solvers.

What would settle it

Run the same token budget with reverse ranking, updating the tokens with the smallest relative noise magnitude each step; if image quality matches the full model as closely as the proposed ranking does, then the noise-magnitude criterion is not the reason CAT Pruning works. Alternatively, measure the Pearson correlation between $n_t - n_{t_0}$ and $n_{t+1} - n_{t_0}$ over all steps and prompts; if the correlation decays below a useful level before the 28th or 50th step, the predictive premise fails.

Watch

Extended reading notes

Core claim

The central discovery is that token-level pruning can be driven by relative noise magnitude, defined as the difference $n_t - n_{t_0}$ between the noise predicted at the current step and at a reference step, and that this one-time ranking remains a reliable guide for later steps. To keep selections consistent across layers and timesteps, the paper clusters tokens once using KMeans on the noise magnitude plus a positional encoding, scores clusters through a light graph pooling layer, and inside each chosen cluster selects tokens by noise magnitude while also injecting stale tokens via an EWMA frequency count. Unselected tokens keep their previous hidden state, so the model recomputes attention and MLP only for the chosen subset. On 1024x1024 generation, this yields a 1.90x speedup at 28 steps and 2.15x at 50 steps on Stable Diffusion v3, with CLIP Score drops of about 0.3 points, and a 1.65x-2.03x speedup on Pixart-Σ with similar quality retention.

Load-bearing premise

The method depends on the assumption that relative noise magnitude at one step reliably predicts which tokens will need updating in later steps, and that reusing frozen hidden states for the other tokens across all layers does not accumulate visible error over dozens of steps.

Editorial extensions

If this is right

  • At 28 sampling steps, Stable Diffusion v3 drops from 168.28T to 90.28T MACs on PartiPrompts, a 1.90x speedup, with CLIP Score moving from 32.33 to 32.03.
  • At 50 steps, the same model drops from 300.50T to 136.70T MACs, a 2.15x speedup, with CLIP Score moving from 32.92 to 32.72.
  • The method combines with DeepCache to reach 2.03x speedup on ImageNet while raising FID from 71.94 to 72.49, a change of 0.55.
  • Because token ranking does not rely on attention maps, CAT Pruning is compatible with online-softmax attention kernels and other intra-kernel acceleration techniques.
  • Spatial clustering plus staleness balancing removes the artifacts seen when tokens are selected purely by noise magnitude, such as stale backgrounds and missing details like windows or eyes.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The same select-by-change and reuse-the-rest principle could transfer to video diffusion and autoregressive image generation, where consecutive states also change slowly and spatially coherently; the paper does not test those settings.
  • Clustering is computed only once at step $t_0+1$; an adaptive or periodic reclustering might improve robustness for prompts where the dominant spatial structure shifts mid-sampling, at some extra cost.
  • The 0.3-point CLIP Score gap suggests the remaining quality loss is concentrated in fine details; pairing token pruning with a lightweight refinement of the unselected tokens, or a stronger staleness scheduler, is a natural next step that the paper leaves open.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper proposes CAT Pruning, a training-free token-pruning and feature-caching method for text-to-image diffusion transformers. At each denoising step, a subset of tokens is updated based on a relative noise magnitude criterion, an EWMA-based staleness term, and a k-means spatial clustering step; the remaining tokens reuse hidden states from the previous step. The method is evaluated on Stable Diffusion v3 and Pixart-Σ using PartiPrompts, COCO2017, and ImageNet, reporting 50% MACs reduction at 28 steps and 60% MACs reduction at 50 steps with end-to-end speedups of 1.65x to 2.15x. The paper also includes an ablation of sparsity levels and a combination with DeepCache.

Significance. If the quality-preservation claim is established, the method would be a useful and practical acceleration technique: it operates without training or fine-tuning, the MACs reductions are large and internally consistent with the reported throughput numbers, the code is promised to be released, and combination with block-level caching is demonstrated. However, the quantitative evidence for preserved quality is currently weaker than the speedup evidence, because the central reported quality metrics lack uncertainty quantification and one configuration shows a CLIP Score drop larger than the paper's 'comparable' claim supports.

major comments (4)
  1. [Table 2, Sec. 4.2] The quality-preservation claim is not supported by the reported numbers for Pixart-Σ on COCO2017. At 28 steps the CLIP Score drops from 31.36 to 30.02, a 1.34-point decrease, and at 50 steps from 31.20 to 30.62, a 0.58-point decrease; these are substantially larger than the drops on the other configurations and are not within the 'comparable' range claimed in the abstract. Since no error bars, per-seed statistics, or number of seeds are reported, it is impossible to tell whether this is noise or a systematic failure of the method on this model/dataset pair.
  2. [Sec. 4.2, Fig. 9] The unpruned-token ratio α = 0.3 is selected by visual inspection on the same evaluation prompts that are later used to report the main results. This is effectively tuning on the test set, so the reported CLIP Scores and FID numbers are optimistic estimates of performance on new prompts. The authors should either select α on a held-out prompt set, report sensitivity to α for the quantitative metrics, or both.
  3. [Sec. 4.1, 'Baselines'] The AT-EDM baseline is not the original AT-EDM method: it combines AT-EDM's token selection with the authors' cache-and-reuse mechanism under a modified token budget, with the stated reason that AT-EDM's similarity-based copy is unsuitable at 30% budget. This may be a reasonable adaptation, but it changes the baseline's algorithmic behavior, and the very low CLIP Scores of the adapted AT-EDM (e.g., 14.66 and 11.00 on COCO2017 in Table 2) suggest the comparison is not representative of AT-EDM as published. The authors should report the original AT-EDM results where available, or justify with ablations that the adaptation is not responsible for the large gap.
  4. [Appendix A.1, Proposition 1] The proof of Proposition 1 is circular as written: it assumes that tokens selected at step t-1 have larger noise magnitudes than unselected tokens, then concludes that selecting tokens by relative noise magnitude is justified, and the claimed predictive relationship between the current step's noise and subsequent steps' noise is never established. The step from 'selected tokens have larger hidden-state changes' to 'tokens with larger current relative noise have larger future relative noise' is exactly the claim that needs empirical or theoretical support. The scatter plot in Figure 3 reports a Pearson correlation but no confidence intervals or per-prompt breakdown, so it does not resolve the concern.
minor comments (4)
  1. [Abstract] The sentence 'Our experiments demonstrate reveal a 50%-60% reduction...' contains a typo; it should read 'demonstrate a 50%-60% reduction' or 'reveal a 50%-60% reduction'.
  2. [Sec. 3.2] The phrase 'the simpliest case as for time-step' and the wording around Proposition 1 are unclear; please rewrite for precision.
  3. [Sec. 4.2] The text 'CAT Pruning consistently achieves better speedup than AT-EDM' is supported by the tables, but the stated reason about 'not rely on the attention map' is not directly evidenced by any experiment and should be presented as a hypothesis.
  4. [Table 3] The FID calculation on ImageNet is described only briefly; please specify the number of generated images, the reference statistics used, and whether the same random seeds were used across methods, since FID values can vary substantially with these choices.

Circularity Check

1 steps flagged · score 4.0 of 10

The empirical quality claim is grounded in external CLIP/FID benchmarks, but the paper's only formal justification for the selection rule, Proposition 1, is circular: its appendix proof assumes the selection criterion and then restates it.

  1. self definitional [Sec. 3.2, Proposition 1; Appendix A.1 (Proof of Proposition 1)]
    "Proposition 1. Selecting tokens with larger relative noise in the current step increases the likelihood that these tokens will exhibit a larger relative noise in subsequent steps. ... At step t−1, we assume: n[i], i ∈ Ts,t−1 > n[i], i ∈ Tu,t−1 ... As a result, at each step, we select tokens based on their relative noise magnitude."

    The proof's opening assumption is exactly the selection rule the proposition is supposed to justify. Because Algorithm 1 copies hidden states for unselected tokens and updates only selected tokens, the inequality MSE(ht, ht−1)[i] for i∈Ts,t > ... i∈Tu,t is true by construction, not by any predictive property of relative noise. The final sentence 'As a result, at each step, we select tokens based on their relative noise magnitude' restates the assumed selection criterion rather than proving that current large relative noise predicts large relative noise in subsequent steps for the same tokens. The proposition therefore provides no independent support for the method; it reduces to its own input.

full rationale

The paper's central speed/quality claim is not circular: the reported MACs reductions follow arithmetically from the 70% token budget (α=0.3), while the quality half is checked against external CLIP Score and FID benchmarks on PartiPrompts, COCO2017, and ImageNet, with comparisons to full models and an AT-EDM baseline. Those numbers are measured, not derived from the method's equations. The one genuinely circular element is the formal justification in Prop. 1/Appendix A.1: the proof assumes the selection rule (selected tokens have larger relative noise), observes by construction that only updated tokens change hidden state, and then concludes that selecting by relative noise is valid. That is a supporting argument whose conclusion is already contained in its premise, so it raises the score to 4. I did not count the choice α=0.3 by visual inspection on evaluation-like prompts as a circular step, because it is a hyperparameter-selection / test-set-tuning concern rather than an equation-level reduction; it is a correctness risk but not a circularity under the stated rules. No load-bearing self-citation or imported uniqueness theorem appears in the paper.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The method's quality claim is empirical and depends on several hand-set hyperparameters (alpha=0.3, t0=8, cluster count=20, EWMA decay, top-k thresholds). The proof of Proposition 1 is an assumption, not a derivation; no invented entities are introduced.

free parameters (5)
  • unpruned token ratio alpha = 0.3
    Chosen in Sec. 4.2 because alpha=0.2 visibly degrades quality; controls the speed-quality trade-off and was not predicted from first principles.
  • pruning start step t0 = 8
    Set in Sec. 4.2, Figure 6, and the DeepCache combination; no ablation is shown for this choice.
  • number of clusters n = 20
    Used in KMeans in Sec. 3.4; chosen by hand, with no sensitivity analysis reported.
  • EWMA decay factor a = not reported
    Appears in Eq. 3.2 for selection-frequency staleness; the value is not given in the paper, yet it controls the balance mechanism.
  • top-cluster and per-cluster top-k counts = not reported
    Algorithm 2 lines 7-16 use topk thresholds that are not specified; these determine how many tokens actually get updated.
assumptions (4)
  • domain assumption Relative noise magnitude n_t - n_{t0} predicts which tokens will need updating at later timesteps.
    Invoked in Sec. 3.2 and Proposition 1; the appendix proof assumes the conclusion and Figure 3 does not report the correlation value in the text.
  • domain assumption Hidden states of unselected tokens can be reused from the previous step without materially changing the final image.
    Used throughout Algorithm 1 (lines 6-8) and Sec. 3.1; inherited from cache-and-reuse literature but not verified with error analysis.
  • domain assumption KMeans with positional encoding yields spatially coherent clusters whose pooled scores identify regions that need updates.
    Introduced in Sec. 3.4 and Algorithm 2; no quantitative validation of cluster quality beyond qualitative images.
  • domain assumption CLIP Score and FID computed on the selected prompts are sufficient proxies for comparable generative quality.
    Used in Sec. 4.2 and 4.3; no user studies or downstream task metrics, and no error bars are reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CAT Pruning: Cluster-Aware Token Pruning For Text-to-Image Diffusion Models." pith.science (2026). https://pith.science/paper/R4RBLQFE

@misc{pith2026250200433,
  author       = {Pith},
  title        = {Pith review of: CAT Pruning: Cluster-Aware Token Pruning For Text-to-Image Diffusion Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/R4RBLQFE}},
  note         = {Machine review of arXiv:2502.00433}
}
read the original abstract

Diffusion models have revolutionized generative tasks, especially in the domain of text-to-image synthesis; however, their iterative denoising process demands substantial computational resources. In this paper, we present a novel acceleration strategy that integrates token-level pruning with caching techniques to tackle this computational challenge. By employing noise relative magnitude, we identify significant token changes across denoising iterations. Additionally, we enhance token selection by incorporating spatial clustering and ensuring distributional balance. Our experiments demonstrate reveal a 50%-60% reduction in computational costs while preserving the performance of the model, thereby markedly increasing the efficiency of diffusion models. The code is available at https://github.com/ada-cheng/CAT-Pruning

Figures

Figures reproduced from arXiv: 2502.00433 by the authors.

Figure 1
Figure 1. CAT Pruning in Stable Diffusion v3. The top row depicts the standard denoising process of Stable Diffusion v3 over 28 inference steps, representing the baseline configuration. The bottom row demonstrates the generative performance of CAT Pruning, which achieves similar generative quality while reducing computation cost by 2× and end-to-end inference time by 1.90×. To carefully ”bypassing” certain patches during the … view at source ↗
Figure 2
Figure 2. Method Overview. At each iteration, tokens are dynamically selected using a combination of the clustering results, noise magnitude, and token staleness. Each part is elaborated in Sec 3.2, Sec 3.3, and Sec 3.4. It is worth noting that we perform clustering only once at step t0 + 1 to avoid computational overhead. 3.1 Token Pruning via Masking Notation Description h Hidden states Ts,t Tokens selected at the iteration… view at source ↗
Figure 3
Figure 3. Scatter plot showing the norm of the relative noise at the current step versus the norm of the [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Visualization of Results Based on Noise Magnitude alone. Selecting tokens purely by noise magnitude causes the indices to center around the teddy bear’s body (as shown in the first row), resulting in noticeable noise artifacts (second row) in the background and a lack …
Figure 5
Figure 5. Figure 5: Visualization of Results Based on Noise Magnitude and Token Staleness. Incorporating both staleness and noise magnitude in token selection yields a more balanced selection distribution, resulting in improved outputs with notably smoother backgrounds and smoother predic…
Figure 6
Figure 6. Figure 6: Comparative Analysis of Token Selection Strategies. The first colomn displays images generated by selecting tokens based on noise magnitude and distributional balance. The second colomn incorporates clustering information for enhanced spatial coherence. The third colom…
Figure 7
Figure 7. Figure 7: The clustering results of different prompts. For each token, clustering is performed based on its relative noise magnitude with positional encoding. We use the K-means algorithm with L2 distance as the clustering metric. 9 [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: Qualitative Results with different sparsity and different prompts. In these cases, even α = 0.2 gives strong results. Prompt : beautiful anime artwork, a cute anime catgirl that looks depressed holding a piece of paper with a smile drawn on it over her mouth, she is ab…
Figure 9
Figure 9. Figure 9: Qualitative Results with different sparsity and different prompts. We find α = 0.3 a sweet spot for the tradeoff between computation efficiency as well as the image quality. 11 [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

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

  1. Forecast then Calibrate: Feature Caching as ODE for Efficient Diffusion Transformers

    cs.CV 2025-08 conditional novelty 6.0 of 10

    A training-free predictor-corrector method that accelerates Diffusion Transformers by solving a feature-ODE, achieving large compute reductions with modest quality loss.

  2. Playing with Transformer at 30+ FPS via Next-Frame Diffusion

    cs.CV 2025-06 conditional novelty 6.0 of 10

    Next-Frame Diffusion combines block-wise causal attention, consistency distillation, and action-based speculative sampling to generate action-conditioned Minecraft video at over 30 FPS on an A100 with a 310M parameter model.

Reference graph

Works this paper leans on

60 extracted references · 25 canonical work pages · cited by 2 Pith papers

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address author booktitle chapter edition editor howpublished institution journal key month note number organization pages publisher school series title type url volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.all := #1 'mid.sentence := ...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    , Cesa-Bianchi, N

    Auer, P. , Cesa-Bianchi, N. and Fischer, P. (2002). Finite-time analysis of the multiarmed bandit problem. Machine Learning, 47 235--256. ://api.semanticscholar.org/CorpusID:207609497

    work page 2002

  4. [4]

    Bianchi, F. M. , Grattarola, D. and Alippi, C. (2020). Spectral clustering with graph neural networks for graph pooling. In Proceedings of the 37th international conference on Machine learning. ACM

    work page 2020

  5. [5]

    Bottou, L. (2010). Large-scale machine learning with stochastic gradient descent. In International Conference on Computational Statistics. ://api.semanticscholar.org/CorpusID:115963355

    work page 2010

  6. [6]

    , Peebles, B

    Brooks, T. , Peebles, B. , Holmes, C. , DePue, W. , Guo, Y. , Jing, L. , Schnurr, D. , Taylor, J. , Luhman, T. , Luhman, E. , Ng, C. , Wang, R. and Ramesh, A. (2024). Video generation models as world simulators. ://openai.com/research/video-generation-models-as-world-simulators

    work page 2024

  7. [7]

    , Monso, D

    Chen, B. , Monso, D. M. , Du, Y. , Simchowitz, M. , Tedrake, R. and Sitzmann, V. (2024 a ). Diffusion forcing: Next-token prediction meets full-sequence diffusion. ://arxiv.org/abs/2407.01392

    arXiv 2024

  8. [8]

    Chen, J. , Ge, C. , Xie, E. , Wu, Y. , Yao, L. , Ren, X. , Wang, Z. , Luo, P. , Lu, H. and Li, Z. (2024 b ). Pixart- : Weak-to-strong training of diffusion transformer for 4k text-to-image generation

    work page 2024

Show all 60 references
  1. [9]

    Chen, J. , Wu, Y. , Luo, S. , Xie, E. , Paul, S. , Luo, P. , Zhao, H. and Li, Z. (2024 c ). Pixart- : Fast and controllable image generation with latent consistency models

    2024

  2. [10]

    Chen, J. , Yu, J. , Ge, C. , Yao, L. , Xie, E. , Wu, Y. , Wang, Z. , Kwok, J. , Luo, P. , Lu, H. and Li, Z. (2023). Pixart- : Fast training of diffusion transformer for photorealistic text-to-image synthesis

    2023

  3. [11]

    Dao, T. , Fu, D. , Ermon, S. , Rudra, A. and R\' e , C. (2022). Flashattention: Fast and memory-efficient exact attention with io-awareness. In Advances in Neural Information Processing Systems (S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho and A. Oh, eds.), vol. 35. ...

    2022

  4. [12]

    , Corrado, G

    Dean, J. , Corrado, G. S. , Monga, R. , Chen, K. , Devin, M. , Le, Q. V. , Mao, M. Z. , Ranzato, M. , Senior, A. W. , Tucker, P. A. , Yang, K. and Ng, A. (2012). Large scale distributed deep networks. In Neural Information Processing Systems. ://api.semanticscholar.org/CorpusID:372467

    2012

  5. [13]

    Decart and Julian Quevedo, S. C. R. W., Quinn McIntyre (2024). Oasis: A universe in a transformer. ://oasis-model.github.io/

    2024

  6. [14]

    , Dong, W

    Deng, J. , Dong, W. , Socher, R. , Li, L.-J. , Li, K. and Fei-Fei, L. (2009). Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition

    2009

  7. [15]

    and Nichol, A

    Dhariwal, P. and Nichol, A. (2021). Diffusion models beat gans on image synthesis. In Advances in Neural Information Processing Systems (M. Ranzato, A. Beygelzimer, Y. Dauphin, P. Liang and J. W. Vaughan, eds.), vol. 34. Curran Associates, Inc. ://proceedings.neurips.cc/paper_...

    2021

  8. [16]

    , Kulal, S

    Esser, P. , Kulal, S. , Blattmann, A. , Entezari, R. , Müller, J. , Saini, H. , Levi, Y. , Lorenz, D. , Sauer, A. , Boesel, F. , Podell, D. , Dockhorn, T. , English, Z. , Lacey, K. , Goodwin, A. , Marek, Y. and Rombach, R. (2024). Scaling rectified flow transformers for high-r...

    2024 arXiv

  9. [17]

    Fang, G. , Ma, X. and Wang, X. (2023). Structural pruning for diffusion models. ://arxiv.org/abs/2305.10924

    2023 arXiv

  10. [18]

    , Holtzman, A

    Hessel, J. , Holtzman, A. , Forbes, M. , Bras, R. L. and Choi, Y. (2022). Clipscore: A reference-free evaluation metric for image captioning. ://arxiv.org/abs/2104.08718

    2022 arXiv

  11. [19]

    Ho, J. (2022). Classifier-free diffusion guidance. ArXiv, abs/2207.12598. ://api.semanticscholar.org/CorpusID:249145348

    2022 arXiv

  12. [20]

    , Jain, A

    Ho, J. , Jain, A. and Abbeel, P. (2020). Denoising diffusion probabilistic models. ://arxiv.org/abs/2006.11239

    2020 arXiv

  13. [21]

    , Aittala, M

    Karras, T. , Aittala, M. , Aila, T. and Laine, S. (2022). Elucidating the design space of diffusion-based generative models. In Proc. NeurIPS

    2022

  14. [22]

    , Cai, T

    Li, M. , Cai, T. , Cao, J. , Zhang, Q. , Cai, H. , Bai, J. , Jia, Y. , Liu, M.-Y. , Li, K. and Han, S. (2024). Distrifusion: Distributed parallel inference for high-resolution diffusion models. ://arxiv.org/abs/2402.19481

    2024 arXiv

  15. [23]

    , Lin, J

    Li, M. , Lin, J. , Meng, C. , Ermon, S. , Han, S. and Zhu, J.-Y. (2022). Efficient spatially sparse inference for conditional gans and diffusion models. In Advances in Neural Information Processing Systems (S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho and A. Oh, eds....

    2022

  16. [24]

    , Liu, Y

    Li, X. , Liu, Y. , Lian, L. , Yang, H. , Dong, Z. , Kang, D. , Zhang, S. and Keutzer, K. (2023). Q-diffusion: Quantizing diffusion models. ://arxiv.org/abs/2302.04304

    2023 arXiv

  17. [25]

    , Maire, M

    Lin, T.-Y. , Maire, M. , Belongie, S. , Bourdev, L. , Girshick, R. , Hays, J. , Perona, P. , Ramanan, D. , Zitnick, C. L. and Dollár, P. (2015). Microsoft coco: Common objects in context. ://arxiv.org/abs/1405.0312

    2015 arXiv

  18. [26]

    , Chen, R

    Lipman, Y. , Chen, R. T. , Ben-Hamu, H. , Nickel, M. and Le, M. (2022). Flow matching for generative modeling. arXiv preprint arXiv:2210.02747

    2022 arXiv

  19. [27]

    , Liu, S

    Liu, H. , Liu, S. , Zhou, Z. , Xu, M. , Xie, Y. , Han, X. , Pérez, J. C. , Liu, D. , Kahatapitiya, K. , Jia, M. , Wu, J.-C. , He, S. , Xiang, T. , Schmidhuber, J. and Pérez-Rúa, J.-M. (2024 a ). Mardini: Masked autoregressive diffusion for video generation at scale. ://arxiv.o...

    2024 arXiv

  20. [29]

    , Gong, C

    Liu, X. , Gong, C. and Liu, Q. (2022). Flow straight and fast: Learning to generate and transfer data with rectified flow. ://arxiv.org/abs/2209.03003

    2022 arXiv

  21. [30]

    , Zhang, X

    Liu, X. , Zhang, X. , Ma, J. , Peng, J. and Liu, Q. (2024 c ). Instaflow: One step is enough for high-quality diffusion-based text-to-image generation. ://arxiv.org/abs/2309.06380

    2024 arXiv

  22. [31]

    and Song, Y

    Lu, C. and Song, Y. (2024). Simplifying, stabilizing and scaling continuous-time consistency models. ://arxiv.org/abs/2410.11081

    2024 arXiv

  23. [32]

    , Zhou, Y

    Lu, C. , Zhou, Y. , Bao, F. , Chen, J. , Li, C. and Zhu, J. (2022). Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. arXiv preprint arXiv:2206.00927

    2022 arXiv

  24. [33]

    , Fang, G

    Ma, X. , Fang, G. , Mi, M. B. and Wang, X. (2024 a ). Learning-to-cache: Accelerating diffusion transformer via layer caching

    2024

  25. [34]

    , Fang, G

    Ma, X. , Fang, G. and Wang, X. (2024 b ). Deepcache: Accelerating diffusion models for free. In The IEEE/CVF Conference on Computer Vision and Pattern Recognition

    2024

  26. [35]

    MacQueen, J. (1967). Some methods for classification and analysis of multivariate observations. ://api.semanticscholar.org/CorpusID:6278891

    1967

  27. [36]

    and Gimelshein, N

    Milakov, M. and Gimelshein, N. (2018). Online normalizer calculation for softmax. ://arxiv.org/abs/1805.02867

    2018 arXiv

  28. [37]

    and Xie, S

    Peebles, W. and Xie, S. (2022). Scalable diffusion models with transformers. arXiv preprint arXiv:2212.09748

    2022 arXiv

  29. [38]

    , Kim, J

    Radford, A. , Kim, J. W. , Hallacy, C. , Ramesh, A. , Goh, G. , Agarwal, S. , Sastry, G. , Askell, A. , Mishkin, P. , Clark, J. , Krueger, G. and Sutskever, I. (2021). Learning transferable visual models from natural language supervision. In International Conference on Machine...

    2021

  30. [39]

    , Fischer, P

    Ronneberger, O. , Fischer, P. and Brox, T. (2015). U-net: Convolutional networks for biomedical image segmentation. ://arxiv.org/abs/1505.04597

    2015 arXiv

  31. [40]

    , Ding, T

    Selvaraju, P. , Ding, T. , Chen, T. , Zharkov, I. and Liang, L. (2024). Fora: Fast-forward caching in diffusion transformer acceleration. ArXiv, abs/2407.01425. ://api.semanticscholar.org/CorpusID:270870209

    2024 arXiv

  32. [41]

    and Malik, J

    Shi, J. and Malik, J. (1997). Normalized cuts and image segmentation. Proceedings of IEEE Computer Society Conference on Computer Vision and Pattern Recognition 731--737. ://api.semanticscholar.org/CorpusID:14848918

    1997

  33. [42]

    , Belkhale, S

    Shih, A. , Belkhale, S. , Ermon, S. , Sadigh, D. and Anari, N. (2023). Parallel sampling of diffusion models

    2023

  34. [43]

    and Dhariwal, P

    Song, Y. and Dhariwal, P. (2023). Improved techniques for training consistency models. ://arxiv.org/abs/2310.14189

    2023 arXiv

  35. [44]

    , Dhariwal, P

    Song, Y. , Dhariwal, P. , Chen, M. and Sutskever, I. (2023). Consistency models. arXiv preprint arXiv:2303.01469

    2023 arXiv

  36. [45]

    and Ermon, S

    Song, Y. and Ermon, S. (2019). Generative modeling by estimating gradients of the data distribution. In Advances in Neural Information Processing Systems (H. Wallach, H. Larochelle, A. Beygelzimer, F. d Alch\' e -Buc, E. Fox and R. Garnett, eds.), vol. 32. Curran Associates, I...

    2019

  37. [46]

    , Sohl-Dickstein, J

    Song, Y. , Sohl-Dickstein, J. , Kingma, D. P. , Kumar, A. , Ermon, S. and Poole, B. (2021). Score-based generative modeling through stochastic differential equations. ://arxiv.org/abs/2011.13456

    2021 arXiv

  38. [47]

    Sutton, R. S. and Barto, A. G. (1998). Reinforcement learning: An introduction. IEEE Trans. Neural Networks, 9 1054--1054. ://api.semanticscholar.org/CorpusID:60035920

    1998

  39. [48]

    , Fatras, K

    Tong, A. , Fatras, K. , Malkin, N. , Huguet, G. , Zhang, Y. , Rector-Brooks, J. , Wolf, G. and Bengio, Y. (2024). Improving and generalizing flow-based generative models with minibatch optimal transport. ://arxiv.org/abs/2302.00482

    2024 arXiv

  40. [49]

    , Cai, Y

    Wang, G. , Cai, Y. , Li, L. , Peng, W. and Su, S. (2024 a ). Pfdiff: Training-free acceleration of diffusion models through the gradient guidance of past and future. ://arxiv.org/abs/2408.08822

    2024 arXiv

  41. [50]

    , Liu, D

    Wang, H. , Liu, D. , Kang, Y. , Li, Y. , Lin, Z. , Jha, N. K. and Liu, Y. (2024 b ). Attention-driven training-free efficiency enhancement of diffusion models. arXiv preprint arXiv:2405.05252

    2024 arXiv

  42. [51]

    , Fang, J

    Wang, J. , Fang, J. , Li, A. and Yang, P. (2024 c ). Pipefusion: Displaced patch pipeline parallelism for inference of diffusion transformer models. ://arxiv.org/abs/2405.14430

    2024 arXiv

  43. [52]

    , Chen, J

    Wang, K. , Chen, J. , Li, H. , Mi, Z. and Zhu, J. (2024 d ). Sparsedm: Toward sparse efficient diffusion models. ://arxiv.org/abs/2404.10445

    2024 arXiv

  44. [53]

    Wimbauer, F. , Wu, B. , Schoenfeld, E. , Dai, X. , Hou, J. , He, Z. , Sanakoyeu, A. , Zhang, P. , Tsai, S. , Kohler, J. et al. (2023). Cache me if you can: Accelerating diffusion models through block caching. arXiv preprint arXiv:2312.03209

    2023 arXiv

  45. [54]

    , Gharbi, M

    Yin, T. , Gharbi, M. , Zhang, R. , Shechtman, E. , Durand, F. , Freeman, W. T. and Park, T. (2024). One-step diffusion with distribution matching distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    2024

  46. [55]

    Yu, J. , Xu, Y. , Koh, J. Y. , Luong, T. , Baid, G. , Wang, Z. , Vasudevan, V. , Ku, A. , Yang, Y. , Ayan, B. K. , Hutchinson, B. , Han, W. , Parekh, Z. , Li, X. , Zhang, H. , Baldridge, J. and Wu, Y. (2022). Scaling autoregressive models for content-rich text-to-image generat...

    2022 arXiv

  47. [56]

    , Gupta, S

    Zhang, W. , Gupta, S. , Lian, X. and Liu, J. (2015). Staleness-aware async-sgd for distributed deep learning. ArXiv, abs/1511.05950. ://api.semanticscholar.org/CorpusID:993719

    2015 arXiv

  48. [57]

    , Liu, H

    Zhang, W. , Liu, H. , Xie, J. , Faccio, F. , Shou, M. Z. and Schmidhuber, J. (2024). Cross-attention makes inference cumbersome in text-to-image diffusion models. arXiv preprint arXiv:2404.02747v1

    2024 arXiv

  49. [58]

    , Fang, T

    Zhao, T. , Fang, T. , Liu, E. , Wan, R. , Soedarmadji, W. , Li, S. , Lin, Z. , Dai, G. , Yan, S. , Yang, H. , Ning, X. and Wang, Y. (2024). Vidit-q: Efficient and accurate quantization of diffusion transformers for image and video generation. ://arxiv.org/abs/2406.02540

    2024 arXiv

  50. [59]

    Zheng, K. , Lu, C. , Chen, J. and Zhu, J. (2023). Dpm-solver-v3: Improved diffusion ode solver with empirical model statistics. ://arxiv.org/abs/2310.13268

    2023 arXiv

  51. [60]

    , Meng, Q

    Zheng, S. , Meng, Q. , Wang, T. , Chen, W. , Yu, N. , Ma, Z. and Liu, T.-Y. (2016). Asynchronous stochastic gradient descent with delay compensation. ArXiv, abs/1609.08326. ://api.semanticscholar.org/CorpusID:3713670

    2016 arXiv

  52. [61]

    , Liu, X

    Zou, C. , Liu, X. , Liu, T. , Huang, S. and Zhang, L. (2024). Accelerating diffusion transformers with token-wise feature caching. ://arxiv.org/abs/2410.05317

    2024 arXiv

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.