Pith. sign in

REVIEW 3 major objections 6 minor 56 references

BRACE: Taming Sharp Irregularities via Barycentric Rational Forecasting for Fast Diffusion Transformers Inference

T0 review · 3 major / 6 minor · reviewed 2026-08-11 · deepseek-v4-flash

Pith's one-line read BRACE replaces derivative-based polynomial extrapolation with a self-normalizing barycentric rational function for diffusion transformer feature caching, and reports superior quality at 3.5x to 5.5x speedups.

desk verdict BRACE's central claim is that rational extrapolation tames sharp DiT features, but the main ImageNet experiments use a configuration that reduces to quadratic Lagrange interpolation, so the paper's flagship results don't test its own mechanism. read the letter →

arxiv 2608.07572 v1 pith:VAHBI7D2 submitted 2026-08-04 cs.CV cs.AI

classification cs.CVcs.AI
keywords DiffusionTransformersfeaturecachingbarycentricrationalinterpolationextrapolationChebyshevweightstraining-freeaccelerationmodelinferenceimageandvideogeneration
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

Diffusion transformers produce high-fidelity images and video but require many sequential network evaluations, and existing cache-then-forecast accelerators predict skipped features with derivative-based polynomial extrapolations that diverge when feature trajectories turn sharply. BRACE is a training-free alternative: it keeps a small sliding window of cached features and predicts the next feature with a barycentric rational function that aggregates raw features directly, using adapted Chebyshev weights instead of finite-difference derivatives. The paper argues that the rational function's self-normalizing denominator absorbs local non-smoothness, so extrapolation error stays bounded by the skip interval and the feature's Lipschitz constant rather than by high-order derivatives. Across DiT-XL/2, FLUX.1-dev, and HunyuanVideo, BRACE reports the best quality–efficiency trade-offs among compared methods at 3.5x to 5.5x speedups, with negligible added computation.

What carries the argument

The central object is the barycentric rational extrapolator in its second form: for cached states $(F_{\tau_j}, x(\tau_j))$, the prediction at target $x(t_{\mathrm{pred}})$ is $F_{\mathrm{pred}} = \left(\sum_j w_j F_{\tau_j}/(x(t_{\mathrm{pred}})-x(\tau_j))\right)/\left(\sum_j w_j/(x(t_{\mathrm{pred}})-x(\tau_j))\right)$, with adapted Chebyshev weights $w_j = (-1)^{j+1}\delta_j$ and $\delta_j \in \{\frac12, 1, \gamma\}$ for the first, interior, and last nodes. An affine domain mapping sends cached timesteps to $[-1,1]$; a FIFO sliding window of capacity $C\le3$ keeps the extrapolation local and makes the Chebyshev nodes coincide with equidistant points. The mechanism doing the work is the self-normalizing denominator $D(x)=\sum_j w_j/(x-x_j)$: instead of estimating derivatives, the predictor is a rational blend of raw features, and the paper's Eq. (11) shows the error is controlled by the skip interval $k$ and the Lipschitz constant $L$ whenever the denominator stays bounded below by a positive structural constant $\lambda$. That denominator is what absorbs sharp irregularities and prevents polynomial overshoot.

What would settle it

Evaluate the denominator $D(\sigma)=\sum_{i=1}^{C} w_i/(\sigma+C-i)$ at the exact weights used in the paper ($C=2$ with $\gamma=0.7$ and $C=3$ with $\gamma=0.5$) over a fine grid of $\sigma\in(0,1]$; if $|D(\sigma)|$ approaches zero anywhere, the bound in Eq. (11) cannot hold.

Watch

Extended reading notes

Core claim

DiT feature trajectories, though globally smooth, are punctuated by localized sharp irregularities, and the paper claims that rational extrapolation is structurally better suited than polynomial extrapolation for forecasting across those transitions. BRACE predicts the next feature map with the second barycentric form $$F_{\mathrm{pred}}=\frac{\sum_{j=1}^{C} w_j\, F_{\tau_j}/(x(t_{\mathrm{pred}})-x(\tau_j))}{\sum_{j=1}^{C} w_j /(x(t_{\mathrm{pred}})-x(\tau_j))},$$ with adapted Chebyshev weights $w_j=(-1)^{j+1}\delta_j$, $\delta_j\in\{\tfrac12,1,\gamma\}$ for first, interior, and last nodes. No derivatives are estimated: the predictor is a weighted rational blend of raw cached features, and the denominator self-normalizes. The stability analysis (Eq. 11) bounds the pointwise error by $(\sum_i |w_i|\, L/\lambda)\, k$, linear in the skip interval $k$ and the Lipschitz constant $L$, in contrast to Taylor extrapolation's remainder $O(s^{m+1}\|F^{(m+1)}\|_\infty)$, which explodes when high-order derivatives spike at sharp turns. The paper reports that this rational predictor outperforms TaylorSeer, HiCache, and FORA on DiT-XL/2, FLUX.1-dev, and HunyuanVideo at matched 3.5x–5.5x speedups.

Load-bearing premise

The stability guarantee rests on two premises: the feature trajectory is Lipschitz continuous, and a structural constant $\lambda>0$ keeps the barycentric denominator bounded away from zero for every skip ratio in $(0,1]$; the proof of the latter is deferred to a supplement not included in the preprint.

Editorial extensions

If this is right

  • At matched FLOPs, BRACE achieves lower FID and sFID than TaylorSeer and HiCache on ImageNet with DiT-XL/2 across 3.56x to 4.98x speedups.
  • On FLUX.1-dev text-to-image at up to 5.55x speedup, BRACE keeps ImageReward and CLIP scores near the full 50-step baseline while reuse- and polynomial-based baselines drop.
  • On HunyuanVideo at 5.56x speedup, BRACE reaches a VBench score of 80.14, the closest to the 80.68 of the unaccelerated 50-step model.
  • Because forecasting uses only a FIFO of cached features and an affine mapping, the added compute is negligible and no retraining is required.
  • The error bound in Eq. (11) predicts that BRACE's degradation grows linearly with the skip interval rather than with high-order derivatives, which is the formal reason it can sustain aggressive skip intervals.

Reading between the lines

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

  • In the headline DiT-XL/2 configuration ($C=3$, $\gamma=0.5$), the adapted Chebyshev weights coincide exactly with Lagrange weights, so the barycentric formula reduces to quadratic polynomial interpolation; the rational structure is therefore exercised mainly in the $C=2$ FLUX and HunyuanVideo settings.
  • The unproved $\lambda>0$ constant is the hinge of the theoretical stability argument, so the supplement's proof of that bound is the first thing to check before the claimed error guarantee can be accepted.
  • The same barycentric forecast recipe could be tested on other sequence-prediction problems with smooth-but-spiky trajectories, such as video frame extrapolation or ODE trajectory forecasting, where derivative-based extrapolators face the same instability.
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

3 major / 6 minor

Summary. The paper presents BRACE, a training-free feature-caching method for accelerating Diffusion Transformer inference. Instead of derivative-based polynomial extrapolation (TaylorSeer, HiCache), BRACE proposes to forecast future hidden states by direct barycentric rational aggregation of a small FIFO cache of historical features, using an adapted Chebyshev weighting scheme. The method is evaluated on DiT-XL/2 for class-conditional ImageNet generation, FLUX.1-dev for text-to-image, and HunyuanVideo for text-to-video, reporting quality-efficiency trade-offs that are claimed to be state of the art. The paper also provides an empirical analysis of feature trajectories and a theoretical stability bound on the extrapolation error.

Significance. If the claims are substantiated, BRACE would be a useful plug-and-play acceleration module with a fresh numerical-analysis flavor, and the cross-architecture evaluation (image and video) is broader than in many caching papers. The idea of replacing finite-difference derivatives with feature-driven rational aggregation is attractive, and the reported latency gains are consistent with the caching paradigm. However, the paper's central mechanism is not actually exercised by its headline ImageNet experiments (the chosen configuration degenerates to a quadratic polynomial), the key stability proof is deferred to a missing supplement, and the empirical reporting lacks code, error bars, and an explanation of implausible metric values. These issues currently limit the paper's significance.

major comments (3)
  1. [Sections 3.3.3-3.3.4, Table 1] The flagship DiT-XL/2 benchmark (Table 1) uses C=3 and gamma=0.5. With these values, Eq. (8) gives weights [1/2, -1, 1/2], which are exactly the barycentric weights of Lagrange interpolation at three equally spaced nodes. Consequently, Eq. (7) simplifies algebraically to the unique quadratic polynomial through the three cached features; the denominator is a removable artifact and the predictor is not a genuine rational function. Thus every BRACE row in Table 1 tests a feature-driven quadratic extrapolator, not the barycentric rational mechanism that the paper's motivation (Section 3.2) and stability analysis (Section 3.4) rely on. The observed FID/sFID gains over TaylorSeer and HiCache in Table 1 therefore do not provide evidence for the paper's central claim that rational self-normalization absorbs sharp feature irregularities. This is an internal inconsistency between the tested configuration and the proposed mechanism, and it needs to be resolved: either report DiT results with a non-degenerate rational setting (e.g., C=2, or C=3 with gamma != 0.5), or explicitly reframe the DiT results as a polynomial extrapolation baseline.
  2. [Section 3.4, Eq. (11)] The proof of the key lower bound |D(s)| >= lambda/k for all sigma in (0,1] is deferred to a 'Supplemental Material' that is not included with the arXiv submission. Since this bound is the basis for the claimed O(k * L) stability guarantee, the central theoretical justification is unverifiable from the manuscript. Moreover, the asserted existence of a universal lambda > 0 for arbitrary gamma in Eq. (8) is not obvious and is in fact false for some parameter values: for C=3 and gamma=0.1, the denominator polynomial 0.5*sigma*(sigma+1) - sigma*(sigma+2) + 0.1*(sigma+1)*(sigma+2) changes sign on (0,1], so D(s) has a pole in the extrapolation interval. The paper needs to state the precise conditions on gamma and C under which Eq. (11) holds, and the proof must be included in the manuscript.
  3. [Tables 1-3 and Reproducibility] No code, project-page link, or evaluation protocol details are provided, and no error bars, standard deviations, or multiple-seed results are reported for any of the FID, ImageReward, or VBench numbers. The differences in Table 1 are small (e.g., FID 2.46 vs. 2.51 at 3.56x), so without variance estimates the claimed consistent improvements are not fully convincing. In addition, Table 2 reports ImageReward 1.0021 for BRACE at 4.16x and 0.9919 for TaylorSeer, both higher than the unaccelerated 50-step FLUX reference (0.9898); it is implausible that a caching method that skips most of the network improves semantic alignment over the full model, and the discrepancy suggests a mismatch in evaluation (e.g., different prompts, seeds, or metric versions) that must be clarified.
minor comments (6)
  1. [Algorithm 1, Section 3.3.2] Line 7 of Algorithm 1 notes that the target domain mapping yields |x(t)| > 1, but Eq. (6) maps cached timesteps only to [-1,1]; the text should state explicitly that extrapolation targets lie outside the canonical interval and explain why this is consistent with the stability analysis.
  2. [Table 2] The reference row for FLUX.1-dev 50 steps contains the garbled entry '0.8800∞1.0000 0.0000'; this appears to be a formatting error (likely a misplaced symbol and missing separator) and should be corrected.
  3. [Section 2.2 and Experiments] FOCA and SpeCA are discussed in the related work as recent forecasting-based methods, but they are not included in any experimental comparison; the authors should either add them as baselines or justify their omission.
  4. [Figures 2 and 3] Several text elements in Figures 2 and 3 render as broken glyph sequences such as '/uni00000013/uni00000014...'; the figures need to be regenerated with a correct font embedding.
  5. [Table 5] The ablation of the boundary sensitivity gamma is performed only for FLUX (C=2). Since the DiT-XL/2 default uses C=3 and gamma=0.5, it would be informative to also ablate gamma for C=3 on DiT to show whether non-degenerate rational configurations are competitive in the headline benchmark.
  6. [Section 3.3.4] The claim that for C<=3 the Chebyshev-Lobatto nodes coincide with equidistant points is correct, but then the adapted weights with gamma are no longer the standard Chebyshev-Lobatto weights; the text should acknowledge this and provide a rationale beyond the empirical tuning in Table 5.

Circularity Check

2 steps flagged · score 2.0 of 10

No circular derivation: the forecast is a deterministic formula, baselines are external, and there are no load-bearing self-citations. The main flags are a mechanism/evidence mismatch (the headline DiT-XL/2 configuration reduces Eq. (7) to Lagrange quadratic interpolation) and a hyperparameter (gamma) selected on the same benchmark used for the SOTA claim.

  1. renaming known result [Section 3.3.4, Eq. (8); Table 1 config 'BRACE(N=4,C=3,gamma=0.5)']
    "w_j = (-1)^(j+1) * delta_j, delta_j = 1/2 (j=1), gamma (j=C), 1 otherwise. ... Crucially, in this low-order regime, these optimal nodes are mathematically identical to equidistant points. ... BRACE(N=4,C=3,gamma=0.5)"

    With C=3 and gamma=0.5, the weights become (1/2, -1, 1/2), which are exactly the barycentric Lagrange weights for three equally spaced nodes. Substituting these into Eq. (7) makes the denominator cancel algebraically, leaving the unique quadratic interpolating polynomial through the three cached features; the claimed self-normalizing rational denominator is a removable artifact. Every DiT-XL/2 row in Table 1 uses this configuration, so the headline results are produced by a polynomial extrapolator presented under the name Barycentric Rational Forecasting, and the rational mechanism that motivates the paper and Eq. (11) is not exercised in the primary benchmark.

  2. fitted input called prediction [Section 4.5, Table 5; Table 2 config 'BRACE(N=5,C=2,gamma=0.7)']
    "Reconstruction metrics (PSNR, SSIM) improve monotonically as gamma -> 1.0. However, perceptual quality (ImageReward) peaks at gamma=0.7 and subsequently declines; thus, gamma=0.7 is established as the default."

    The default gamma is selected using the same FLUX.1-dev/DrawBench evaluation metrics (ImageReward, PSNR, SSIM, LPIPS) that Table 2 later reports as BRACE's state-of-the-art result. The hyperparameter is therefore fit to the benchmark whose prediction it supports, so the Table 2 numbers are not a fully out-of-sample validation of gamma. This is a mild fit-to-data issue rather than a forced identity, so it contributes only slightly to the circularity score.

full rationale

The paper's derivation chain is not circular in the core sense: the forecast of future features is a fixed barycentric formula, not a quantity fitted to the final FID/PSNR, and the comparisons against TaylorSeer, HiCache, and FORA are external published baselines. There are no load-bearing self-citations: the numerical-analysis premises cite Berrut-Trefethen, Stoer-Bulirsch, and Webb-Trefethen-Gonnet, independent of the present authors, and no prior work by these authors is invoked to justify the method. The first flagged step is better described as a mechanism-evidence mismatch than as circularity: in the flagship DiT-XL/2 setting, Eq. (8) with C=3 and gamma=0.5 yields the exact Lagrange barycentric weights, so Eq. (7) reduces to quadratic polynomial interpolation; the claimed rational self-normalization is not tested in Table 1, and the 'Chebyshev enhancement' is also vacuous because the paper itself states that the optimal nodes are identical to equidistant points in the low-order regime used everywhere. The second flagged step is a mild benchmark-selection issue: gamma=0.7 is chosen using the same FLUX benchmark that later supports the SOTA claim, but the chosen value is a compromise across metrics and the underlying extrapolator is not derived from the target metric. Additionally, the stability bound in Eq. (11) relies on an unproved lambda>0 lemma deferred to a Supplement not present in the arXiv v1; this is an omitted proof and a correctness risk, not circularity. Overall, the derivation is self-contained and the reported numbers are honest external evaluations, so the appropriate circularity score is 2.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The method introduces no new physical entities. The free parameters gamma and C are tuned per benchmark and affect the central quality claims. The unproven lambda bound and the Lipschitz assumption are the key unverified premises in the stability argument.

free parameters (2)
  • gamma (boundary sensitivity coefficient) = 0.5 for DiT-XL/2, 0.7 for FLUX.1-dev, 0.4 for HunyuanVideo
    Tuned per model to maximize reported metrics; the stability analysis does not specify an optimal value and the ablation shows sensitivity to it.
  • C (window capacity) = 3 for DiT-XL/2, 2 for FLUX.1-dev, 2 for HunyuanVideo
    Chosen per model through ablation; Table 4 shows C=4 degrades performance, and the paper claims C<=3 is needed to align Chebyshev nodes with equidistant points.
assumptions (3)
  • ad hoc to paper The barycentric denominator D(x) is bounded below by lambda > 0 for all s in (0,k] (Eq. 10-11).
    The proof is deferred to a 'Supplemental Material' not included in the preprint; no numerical verification of lambda is provided.
  • domain assumption DiT feature trajectories are Lipschitz continuous with bounded L (Eq. 9).
    The paper assumes |F(x)-F(x_i)| <= L|x-x_i|, but empirically the features exhibit sharp irregularities, so a finite L is not demonstrated.
  • standard math For C<=3, Chebyshev-Gauss-Lobatto nodes coincide with equidistant points, making the adapted weights theoretically optimal.
    This is true for C<=3, but the 'optimality' for extrapolation of DiT features is not established.

how reviews work

0 comments
Cite this review

Pith. "Pith review of BRACE: Taming Sharp Irregularities via Barycentric Rational Forecasting for Fast Diffusion Transformers Inference." pith.science (2026). https://pith.science/paper/VAHBI7D2

@misc{pith2026260807572,
  author       = {Pith},
  title        = {Pith review of: BRACE: Taming Sharp Irregularities via Barycentric Rational Forecasting for Fast Diffusion Transformers Inference},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VAHBI7D2}},
  note         = {Machine review of arXiv:2608.07572}
}
read the original abstract

Diffusion Transformers (DiTs) have demonstrated exceptional performance in high-fidelity image and video generation. To alleviate their massive computational overhead, temporal feature caching has been proposed to bypass redundant computations. However, existing cache-then-forecast methods driven by derivative-based polynomials often cause severe quality degradation under high acceleration due to unstable long-step predictions. To address this bottleneck, we propose Barycentric Rational Forecasting with Chebyshev Enhancement (BRACE). Motivated by the observation that DiT feature trajectories are globally smooth yet frequently exhibit sharp irregularities and local non-smoothness, BRACE shifts the paradigm from derivative-driven polynomial extrapolation to feature-driven rational forecasting. Specifically, it maintains a local sliding window to cache sparse historical features and leverages adapted Chebyshev weights to formulate a barycentric rational function, directly aggregating these raw features to ensure numerical stability. Extensive experiments demonstrate that BRACE achieves state-of-the-art quality-efficiency trade-offs across various DiT architectures with negligible computational overhead.

Figures

Figures reproduced from arXiv: 2608.07572 by the authors.

Figure 1
Figure 1. Feature geometric complexity. (a) Trajectories ex [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. DiT feature dynamics and basis expressiveness. (a) [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Overview of the BRACE framework. Left: The system-level pipeline manages a FIFO Sliding Window to maintain a [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison of different acceleration [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Visualization examples for different acceleration methods on Flux. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Qualitative comparison on Text-to-Video Generation. [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

56 extracted references · 19 canonical work pages

  1. [1]

    Hyojin Bahng, Caroline Chan, Fredo Durand, and Phillip Isola. 2025. Cycle Con- sistency as Reward: Learning Image-Text Alignment without Human Preferences. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). 22934–22946

  2. [2]

    Jean-Paul Berrut and Lloyd N Trefethen. 2004. Barycentric lagrange interpolation. SIAM review46, 3 (2004), 501–517

  3. [3]

    Daniel Bolya and Judy Hoffman. 2023. Token merging for fast stable diffusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 4599–4603

  4. [4]

    Qi Cai, Jingwen Chen, Yang Chen, Yehao Li, Fuchen Long, Yingwei Pan, Zhaofan Qiu, Yiheng Zhang, Fengbin Gao, Peihan Xu, Yimeng Wang, Kai Yu, Wenxuan Chen, Ziwei Feng, Zijian Gong, Jianzhuang Pan, Yi Peng, Rui Tian, Siyu Wang, Bo Zhao, Ting Yao, and Tao Mei. 2025. HiDream-I1: A High- Efficient Image Generative Foundation Model with Sparse Diffusion Transfo...

  5. [5]

    Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhong- dao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. 2023. PixArt-𝛼: Fast Training of Diffusion Transformer for Photorealistic Text-to-Image Synthesis. arXiv:2310.00426 [cs.CV] https://arxiv.org/abs/2310.00426

  6. [6]

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. 2009. Ima- geNet: A large-scale hierarchical image database. In2009 IEEE Conference on Com- puter Vision and Pattern Recognition. 248–255. doi:10.1109/CVPR.2009.5206848

  7. [7]

    Gongfan Fang, Xinyin Ma, and Xinchao Wang. 2023. Structural Pruning for Diffusion Models. arXiv:2305.10924 [cs.LG] https://arxiv.org/abs/2305.10924

  8. [8]

    Liang Feng, Shikang Zheng, Jiacheng Liu, Yuqi Lin, Qinming Zhou, Peiliang Cai, Xinyu Wang, Junjie Chen, Chang Zou, Yue Ma, and Linfeng Zhang. 2026. Hi- Cache: A Plug-in Scaled-Hermite Upgrade for Taylor-Style Cache-then-Forecast Diffusion Acceleration. InInternational Conference on Learning Representations (ICLR)

Show all 56 references
  1. [9]

    Michael S Floater and Kai Hormann. 2007. Barycentric rational interpolation with no poles and high rates of approximation.Numer. Math.107, 2 (2007), 315–331

  2. [10]

    Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi

  3. [11]

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. 2018. GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium. arXiv:1706.08500 [cs.LG] https://arxiv.org/abs/1706. 08500

  4. [12]

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. 2020. Denoising Diffusion Probabilis- tic Models. InAdvances in Neural Information Processing Systems, H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin (Eds.), Vol. 33. Curran Asso- ciates, Inc., 6840–6851. https://proce...

  5. [13]

    Alain Horé and Djemel Ziou. 2010. Image Quality Metrics: PSNR vs. SSIM. In 2010 20th International Conference on Pattern Recognition. 2366–2369. doi:10.1109/ ICPR.2010.579

  6. [14]

    Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuan- han Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. 2024. Vbench: Comprehensive benchmark suite for video generative models. InProceedings of the IEEE/CVF Conference on Computer Vision ...

  7. [15]

    Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, et al. 2024. Hunyuanvideo: A systematic framework for large video generative models.arXiv preprint arXiv:2412.03603 (2024)

  8. [16]

    Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catan- zaro. 2021. DiffWave: A Versatile Diffusion Model for Audio Synthesis. arXiv:2009.09761 [eess.AS] https://arxiv.org/abs/2009.09761

  9. [17]

    Black Forest Labs. 2024. FLUX. https://github.com/black-forest-labs/flux

  10. [18]

    Black Forest Labs, Stephen Batifol, Andreas Blattmann, Frederic Boesel, Saksham Consul, Cyril Diagne, Tim Dockhorn, Jack English, Zion English, Patrick Esser, Sumith Kulal, Kyle Lacey, Yam Levi, Cheng Li, Dominik Lorenz, Jonas Müller, Dustin Podell, Robin Rombach, Harry Saini,...

  11. [19]

    Xiuyu Li, Yijiang Liu, Long Lian, Huanrui Yang, Zhen Dong, Daniel Kang, Shang- hang Zhang, and Kurt Keutzer. 2023. Q-diffusion: Quantizing diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision. 17535– 17545

  12. [20]

    Haohe Liu, Zehua Chen, Yi Yuan, Xinhao Mei, Xubo Liu, Danilo Mandic, Wenwu Wang, and Mark D Plumbley. 2023. Audioldm: Text-to-audio generation with latent diffusion models.arXiv preprint arXiv:2301.12503(2023)

  13. [21]

    Jiacheng Liu, Chang Zou, Yuanhuiyi Lyu, Junjie Chen, and Linfeng Zhang. 2025. From reusing to forecasting: Accelerating diffusion models with taylorseers. In Proceedings of the IEEE/CVF International Conference on Computer Vision. 15853– 15863

  14. [22]

    Jiacheng Liu, Chang Zou, Yuanhuiyi Lyu, Fei Ren, Shaobo Wang, Kaixin Li, and Linfeng Zhang. 2025. Speca: Accelerating diffusion transformers with speculative feature caching. InProceedings of the 33rd ACM International Conference on Multimedia. 10024–10033

  15. [23]

    Xingchao Liu, Chengyue Gong, and Qiang Liu. 2022. Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow. arXiv:2209.03003 [cs.LG] https://arxiv.org/abs/2209.03003

  16. [24]

    Alexander Lobashev, Dmitry Guskov, Maria Larchenko, and Mikhail Tamm

  17. [25]

    Jinming Lou, Wenyang Luo, Yufan Liu, Bing Li, Xinmiao Ding, Weiming Hu, Yuming Li, and Chenguang Ma. 2024. Token caching for diffusion transformer acceleration.arXiv preprint arXiv:2409.18523(2024)

  18. [26]

    Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu

  19. [27]

    Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. 2025. DPM-Solver++: Fast Solver for Guided Sampling of Diffusion Probabilistic Models. Machine Intelligence Research22, 4 (June 2025), 730–751. doi:10.1007/s11633-025- 1562-4

  20. [28]

    Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. 2023. Latent Con- sistency Models: Synthesizing High-Resolution Images with Few-Step Inference. arXiv:2310.04378 [cs.CV] https://arxiv.org/abs/2310.04378

  21. [29]

    Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps.Advances in neural information processing systems35 (2022), 5775–5787

  22. [30]

    Xinyin Ma, Gongfan Fang, and Xinchao Wang. 2024. Deepcache: Accelerating diffusion models for free. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 15762–15772

  23. [31]

    William Peebles and Saining Xie. 2023. Scalable diffusion models with transform- ers. InProceedings of the IEEE/CVF international conference on computer vision. 4195–4205

  24. [32]

    Nanye Ma, Mark Goldstein, Michael S Albergo, Nicholas M Boffi, Eric Vanden- Eijnden, and Saining Xie. 2024. Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers. InEuropean Conference on Com- puter Vision. Springer, 23–40

  25. [33]

    Tim Salimans and Jonathan Ho. 2022. Progressive Distillation for Fast Sampling of Diffusion Models. arXiv:2202.00512 [cs.LG] https://arxiv.org/abs/2202.00512

  26. [34]

    Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. 2023. Adversarial Diffusion Distillation. arXiv:2311.17042 [cs.CV] https://arxiv.org/ abs/2311.17042

  27. [35]

    Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi

    Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Den- ton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi. 2022. Photorealistic Text-to-Image Diffu...

  28. [36]

    Yuzhang Shang, Zhihang Yuan, Bin Xie, Bingzhe Wu, and Yan Yan. 2023. Post- training Quantization on Diffusion Models. InCVPR

  29. [37]

    Jiaming Song, Chenlin Meng, and Stefano Ermon. 2022. Denoising Diffusion Implicit Models. arXiv:2010.02502 [cs.LG] https://arxiv.org/abs/2010.02502

  30. [38]

    Pratheba Selvaraju, Tianyu Ding, Tianyi Chen, Ilya Zharkov, and Luming Liang

  31. [39]

    STOER and R

    J. STOER and R. BULiRSCH. 1966. Numerical Treatment of Ordinary Differential Equations by Extrapolation Methods.Numer. Math.8 (1966), 1–13. http://eudml. org/doc/131682

  32. [40]

    Trefethen and David Bau

    Lloyd N. Trefethen and David Bau. 1997.Numerical Linear Algebra. SIAM

  33. [41]

    Team Wan, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianxiao Yang, Jianyuan Zeng, Jiayu Wang, Jingfeng Zhang, Jingren Zhou, Jinkai Wang, Jixuan Chen, Kai Zhu, Kang Zhao, Keyu Yan, Lianghua Huang, Mengyang Feng, Ningyi Zhang, Pande...

  34. [42]

    Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. 2023. Consistency Models. arXiv:2303.01469 [cs.LG] https://arxiv.org/abs/2303.01469

  35. [43]

    Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. 2023. ImageReward: learning and evaluating human preferences for text-to-image generation. InProceedings of the 37th International Conference on Neural Information Processing System...

  36. [44]

    Zhihang Yuan, Hanling Zhang, Lu Pu, Xuefei Ning, Linfeng Zhang, Tianchen Zhao, Shengen Yan, Guohao Dai, and Yu Wang. 2024. DiTFastAttn: Attention Compression for Diffusion Transformer Models. InThe Thirty-eighth Annual Conference on Neural Information Processing Systems. https...

  37. [45]

    Evelyn Zhang, Bang Xiao, Jiayi Tang, Qianli Ma, Chang Zou, Xuefei Ning, Xuming Hu, and Linfeng Zhang. 2024. Token Pruning for Caching Better: 9 Times Acceleration on Stable Diffusion for Free. arXiv:2501.00375 [cs.CV] https://arxiv. org/abs/2501.00375

  38. [46]

    Marcus Webb, Lloyd N Trefethen, and Pedro Gonnet. 2012. Stability of barycentric interpolation formulas for extrapolation.SIAM Journal on Scientific Computing MM ’26, November 10–14, 2026, Rio de Janeiro, Brazil Yang et al. 34, 6 (2012), A3009–A3015

  39. [47]

    Wenliang Zhao, Lujia Bai, Yongming Rao, Jie Zhou, and Jiwen Lu. 2023. Unipc: A unified predictor-corrector framework for fast sampling of diffusion models. Advances in Neural Information Processing Systems36 (2023), 49842–49869

  40. [48]

    Xuanlei Zhao, Xiaolong Jin, Kai Wang, and Yang You. 2024. Real-time video generation with pyramid attention broadcast.arXiv preprint arXiv:2408.12588 (2024)

  41. [49]

    Shikang Zheng, Liang Feng, Xinyu Wang, Qinming Zhou, Peiliang Cai, Chang Zou, Jiacheng Liu, Yuqi Lin, Junjie Chen, Yue Ma, and Linfeng Zhang. 2026. Fore- cast then Calibrate: Feature Caching as ODE for Efficient Diffusion Transformers. InProceedings of the AAAI Conference on A...

  42. [50]

    Efros, Eli Shechtman, and Oliver Wang

    Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang

  43. [55]

    Chang Zou, Xuyang Liu, Ting Liu, Siteng Huang, and Linfeng Zhang

  44. [56]

    arXiv:2410.05317 [cs.LG] https://arxiv.org/abs/2410.05317

    Accelerating Diffusion Transformers with Token-wise Feature Caching. arXiv:2410.05317 [cs.LG] https://arxiv.org/abs/2410.05317

  45. [2018]

    InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

    The Unreasonable Effectiveness of Deep Features as a Perceptual Metric. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

  46. [2022]

    arXiv:2104.08718 [cs.CV] https://arxiv.org/abs/2104.08718

    CLIPScore: A Reference-free Evaluation Metric for Image Captioning. arXiv:2104.08718 [cs.CV] https://arxiv.org/abs/2104.08718

  47. [2024]

    arXiv:2407.01425 [cs.CV] https://arxiv.org/abs/2407.01425

    FORA: Fast-Forward Caching in Diffusion Transformer Acceleration. arXiv:2407.01425 [cs.CV] https://arxiv.org/abs/2407.01425

  48. [2025]

    Hessian geometry of latent space in generative models.arXiv preprint arXiv:2506.10632(2025)

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.