Pith. sign in

REVIEW 4 major objections 5 minor 4 cited by

FoCa treats feature caching as solving an ODE and reports near-lossless speedups of up to 6.45x on diffusion transformers, with no retraining.

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 ↗

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

T0 review reviewed 2026-08-05 challenge →

load-bearing objection Promising training-free caching method with a genuinely new predictor-corrector combo, but the headline speedups mix FLOPs with wall-clock and the Heun equation as written does not match the method. the 4 major comments →

arxiv 2508.16211 v1 pith:WPF2XANG submitted 2025-08-22 cs.CV

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

Shikang Zheng , Liang Feng , Xinyu Wang , Qinming Zhou , Peiliang Cai , Chang Zou , Jiacheng Liu , Yuqi Lin
show 3 more authors
Junjie Chen Yue Ma Linfeng Zhang
This is my paper
classification cs.CV
keywords feature cachingdiffusion transformersODE solverspredictor-correctorBDF2Heun's methodtraining-free accelerationdiffusion inference
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

The paper argues that the hidden features of diffusion transformers evolve smoothly along the denoising trajectory and can be modeled as a feature-ODE. It claims current caching methods fail because they either copy old features or extrapolate with high-order finite differences, both of which become unstable over long skip intervals. FoCa instead predicts skipped features with a second-order backward differentiation formula, then corrects the prediction with a Heun-style step that blends in the most recent fully computed feature. If correct, this gives a training-free way to run large generative image and video models several times faster while keeping quality nearly intact.

Core claim

The central claim is that feature caching should be viewed as numerically integrating a feature-ODE, dF/dt = g_theta(F, t), along the reverse-time sampling grid. FoCa uses BDF2 to forecast the hidden state at a skipped timestep from the two most recent cached tensors, then applies a Heun corrector that mixes the forecast with the latest fully computed activation to damp overshoot. The paper reports near-lossless speedups of 5.50x on FLUX, 6.45x on HunyuanVideo, 3.17x on Inf-DiT, and 4.53x on DiT-XL/2, with better FID than prior caching baselines at equal FLOP budgets. The main contrast is with TaylorSeer, whose higher-order finite-difference forecasts the paper identifies as numerically unst

What carries the argument

The central object is the feature-ODE, an equation describing the evolution of a layer's hidden representation as a function of denoising time, derived via the chain rule from the probability-flow ODE. The method is a predictor-corrector: a second-order backward differentiation formula aggregates the two most recent cached feature tensors to forecast the next skipped state, and a Heun-style corrector averages the derivative at the latest fully computed step with the derivative at the forecast point to damp error. This lets FoCa use historical features directly without estimating high-order derivatives, which the paper identifies as the source of instability in prior forecasting methods.

Load-bearing premise

The load-bearing premise is that replacing the current-step slope in the Heun corrector with the slope from the most recent fully computed step remains both valid and stable; the paper asserts an error bound for this substitution but defers the proof to an appendix that is not present.

What would settle it

Implement the FoCa corrector using the exact derivative at the predicted point, as in standard Heun's method, and compare feature error and FID against the paper's Eq. 8 on DiT-XL/2 at N=5. If the standard version is no worse and the paper's version shows error growing roughly linearly with skip length, the claimed stability mechanism is not doing the work. Alternatively, measure the maximum feature deviation over many seeds across skip intervals and check whether it stays bounded as Proposition 1 claims.

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

If this is right

  • FoCa can be dropped into pretrained diffusion transformers with no retraining and no architectural changes, so it applies to existing deployed models.
  • At equal FLOP budgets, FoCa reports better FID and ImageReward scores than reuse-based baselines such as FORA, ToCa, DuCa, and TeaCache, and than forecast-based TaylorSeer.
  • The 6.45x speedup on HunyuanVideo keeps the VBench score within about one point of the original sampler, indicating near-lossless video quality.
  • The ablation shows the Heun corrector is load-bearing: BDF2-only forecasting degrades rapidly under large skip intervals, while FoCa stays stable.
  • The feature-ODE perspective reframes caching design as a numerical integration problem, opening the same machinery to other multistep ODE solvers beyond BDF2-Heun.

Where Pith is reading between the lines

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

  • The same predictor-corrector scheme could apply to other sequential inference settings where cached intermediate representations drift smoothly, such as autoregressive decoding or recurrent video generation, though the paper does not test these.
  • The stiffness analysis suggests a fixed cache interval is suboptimal: error concentrates in a late stiff zone, so an adaptive schedule that caches more densely near the end of sampling could push speedups beyond the uniform intervals reported.
  • Because FoCa is training-free and uses only cached tensors, it should stack multiplicatively with orthogonal accelerations such as step reduction (DDIM, DPM-Solver) and model compression, a combination the paper does not evaluate explicitly.
  • The paper's stability guarantee rests on a nonstandard Heun corrector whose proof is deferred to a missing appendix; testing the error bound directly on feature trajectories would settle whether the claimed stability holds beyond the reported checkpoints.
Share X Bluesky LinkedIn Reddit HN

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

4 major / 5 minor

Summary. The manuscript proposes FoCa, a training-free feature-caching method for Diffusion Transformers. It reinterprets the sequence of hidden features across denoising timesteps as a solution to a feature ODE (Eq. 5), then predicts skipped features using a BDF2-style predictor (Eq. 7) followed by a Heun-style corrector (Eq. 8). Experiments on FLUX, HunyuanVideo, Inf-DiT, and DiT-XL/2 report large FLOPs reductions with small or no quality loss, and an ablation attributes stability to the corrector. The paper also includes a stiffness analysis and states Proposition 1 bounding the prediction error independently of skip length, with the proof deferred to a missing Appendix A.1.

Significance. If the reported results hold, FoCa would be a useful training-free accelerator: it applies across four architectures and tasks, preserves generation quality at aggressive caching ratios, and connects feature caching to numerical integration of an ODE. The breadth of evaluation and the simplicity of the predictor-corrector design are strengths. However, the theoretical guarantee is currently unverifiable because the proof is missing and the corrector as written deviates from standard Heun, and the headline 'speedups' are FLOPs reductions rather than measured latency. These issues are load-bearing for the central near-lossless-speedup claim and need to be resolved before the contribution can be assessed.

major comments (4)
  1. [Method, Eq. (8)] The corrector is not the standard Heun (explicit trapezoidal) method. Heun averages f(t_k, y_k) and f(t_{k+1}, y_pred) over [t_k, t_{k+1}]; Eq. (8) instead uses F^(1)(x_{k-N}) as the first slope and states that x_{k-N} is the latest fully computed step. If k is the current step, this slope is N steps stale; if k is the last full computation, the index should be x_k. Either way the formula is inconsistent with the text. Since Fig. 7 shows the corrector is essential for stability and Proposition 1 is the only theoretical support, please correct Eq. (8) or provide a dedicated error analysis for the stale-derivative update actually used.
  2. [Proposition 1 / Appendix A.1] The proof of Proposition 1 is deferred to Appendix A.1, but the submitted manuscript contains no appendix. Equation (9) introduces rho and tau_max without definitions and asserts a bound independent of k without stating the assumptions on F and the corrector. The authors' central robustness claim, and the claimed advantage over TaylorSeer and reuse methods, rests on this result. The proof, constant definitions, and regularity assumptions must be included in the manuscript.
  3. [Experiments, Tables 1-2 and Abstract/Conclusion] The abstract and conclusion report speedups of 5.50x on FLUX and 6.45x on HunyuanVideo, but these correspond to FLOPs reductions, not wall-clock latency. Table 1 FoCa N=7 shows 4.05x latency speedup vs 5.54x FLOPs; Table 2 FoCa N=7 shows 4.64x latency vs 6.45x FLOPs. User-visible speedup is determined by latency; FLOPs counts can overstate gains when memory access, attention overhead, and scheduling costs are significant. Please report measured latency speedups as the headline numbers, or clearly label the headline numbers as FLOPs reductions.
  4. [Analysis on stiffness] The paper claims the BDF2-Heun predictor-corrector is well-suited for stiff ODEs, but the implemented update is explicit (the Heun corrector is an explicit method, and the BDF2 formula is used only to generate a prediction, not as an implicit solver). Figure 8 introduces LTE and a Stiffness Index without defining either or describing how they are computed. As written, this supports the method selection only qualitatively. Please provide formal definitions and either a quantitative stiffness/stability analysis or soften the claim.
minor comments (5)
  1. [Notation, Eqs. (5)-(8)] The indexing of k and h_k is ambiguous: h_k appears in Eqs. (7)-(8) but the relationship between the caching interval N, the step size, and the indices k, k-N is not defined. Please clarify the grid and state whether steps are uniform.
  2. [Tables 1-3] Report standard deviations or confidence intervals for stochastic metrics such as FID. In particular, FoCa N=2 improving FID from 2.32 to 2.17 on ImageNet is likely within noise and should not be emphasized without repeated runs.
  3. [Appendix] The text says 'Please refer to Appendix for details' but the arXiv version has no appendix. Include the missing appendix with experiment settings, hyperparameters, and the proof of Proposition 1.
  4. [Abstract/Table 1] The abstract says 5.50x on FLUX while Table 1 reports 5.54x FLOPs speedup for FoCa N=7. Use consistent rounding and specify which measure is meant.
  5. [Table 4 caption] Typo: 'gene tative' should be 'generative'.

Circularity Check

0 steps flagged

No significant circularity: FoCa is a parameter-free predictor-corrector scheme; the speedups are empirical measurements against independent baselines, and the self-cited TaylorSeer/ToCa are comparisons, not premises.

full rationale

The derivation chain is not circular. Equation (5) introduces a feature-ODE by the chain rule, explicitly noting that g_theta is intractable; the subsequent BDF2 predictor (Eq. 7) and Heun corrector (Eq. 8) operate only on cached feature values and contain no fitted parameters tied to the reported FID, ImageReward, or VBench scores. The headline speedups are measured outcomes, not consequences forced by the ODE formalism. TaylorSeer and ToCa, which involve overlapping authors, are used as baselines in comparisons rather than as justification for FoCa's design, so these self-citations are not load-bearing. The ODE framing is partly definitional, but it does not reduce the prediction to its inputs: the predictor-corrector is a concrete numerical integration scheme whose behavior is ablated empirically in Fig. 7. The main substantive weakness is that Proposition 1's error bound is deferred to Appendix A.1, which is not present in v1; this is an omitted-proof concern, not a circularity, because the bound is claimed as a theorem rather than assumed as an input. Overall, no step in the paper's argument reconstructs its conclusion from its own premises by construction.

Axiom & Free-Parameter Ledger

2 free parameters · 3 axioms · 0 invented entities

The main free parameter is the cache interval N, selected by the user. The method introduces no learned parameters, but the proof of stability relies on undefined constants rho and tau_max. The key domain assumption is that hidden features evolve smoothly enough for multistep ODE solvers, and the key ad-hoc assumption is the replacement of the current slope by the slope from an older timestep in the Heun corrector.

free parameters (2)
  • cache interval N = 5,7,8 on FLUX; 7 on HunyuanVideo; 3,5 on DiT; 5 on Inf-DiT
    Chosen per model and acceleration target; affects speed-quality trade-off.
  • stability constants rho, tau_max in Prop 1
    Introduced in the bound but never defined or estimated.
axioms (3)
  • domain assumption Hidden features along the reverse diffusion trajectory satisfy a smooth ODE with a tractable Lipschitz drift (Eq. 5)
    Assumed to justify applying multistep solvers; no evidence of smoothness or bounded stiffness.
  • ad hoc to paper The future derivative F'(x^{k+1}) in the Heun corrector can be replaced by the derivative at the previous fully-computed step F'(x^{k-N})
    Eq. 8 uses F'(x^{k-N}) rather than F'(x^k), which is not standard Heun and is unproven.
  • standard math Standard chain rule / differentiability of network blocks
    Required for Eq. 5.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Forecast then Calibrate: Feature Caching as ODE for Efficient Diffusion Transformers." pith.science (2026). https://pith.science/paper/WPF2XANG

@misc{pith2026250816211,
  author       = {Pith},
  title        = {Pith review of: Forecast then Calibrate: Feature Caching as ODE for Efficient Diffusion Transformers},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WPF2XANG}},
  note         = {Machine review of arXiv:2508.16211}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Diffusion Transformers (DiTs) have demonstrated exceptional performance in high-fidelity image and video generation. To reduce their substantial computational costs, feature caching techniques have been proposed to accelerate inference by reusing hidden representations from previous timesteps. However, current methods often struggle to maintain generation quality at high acceleration ratios, where prediction errors increase sharply due to the inherent instability of long-step forecasting. In this work, we adopt an ordinary differential equation (ODE) perspective on the hidden-feature sequence, modeling layer representations along the trajectory as a feature-ODE. We attribute the degradation of existing caching strategies to their inability to robustly integrate historical features under large skipping intervals. To address this, we propose FoCa (Forecast-then-Calibrate), which treats feature caching as a feature-ODE solving problem. Extensive experiments on image synthesis, video generation, and super-resolution tasks demonstrate the effectiveness of FoCa, especially under aggressive acceleration. Without additional training, FoCa achieves near-lossless speedups of 5.50 times on FLUX, 6.45 times on HunyuanVideo, 3.17 times on Inf-DiT, and maintains high quality with a 4.53 times speedup on DiT.

Figures

Figures reproduced from arXiv: 2508.16211 by Chang Zou, Jiacheng Liu, Junjie Chen, Liang Feng, Linfeng Zhang, Peiliang Cai, Qinming Zhou, Shikang Zheng, Xinyu Wang, Yue Ma, Yuqi Lin.

Figure 1
Figure 1. Figure 1: (a) PCA visualization of feature trajectory: FoCa [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Visualization of the image generated by different methods on prompt: [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Comparison between previous caching methods [PITH_FULL_IMAGE:figures/full_fig_p002_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The computation of FoCa for xt−1, xt−2, xt−3, respectively. For each skipping step, FoCa predicts a future hidden state using the two most recent timestep, then applies a Heun corrector that blends the prediction with the most recent fully￾computed feature so that enables accurate and stable prediction under large skip intervals. accelerate diffusion inference by leveraging historical fea￾tures for stable … view at source ↗
Figure 5
Figure 5. Figure 5: Visual comparison of 5.5× accelerated FLUX. Results on Text-to-Image Generation. As summarized in [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: Ablation study of Heun Correction. Adding the [PITH_FULL_IMAGE:figures/full_fig_p006_7.png] view at source ↗
Figure 6
Figure 6. Figure 6: Visualization of different caching methods on HunyuanVideo. FoCa maintains high-quality generation under the [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 8
Figure 8. Figure 8: Stiffness Analysis: (a) LTE and Stiffness Index of FLUX across timesteps, revealing a sharp rise in the late stage that indicates a numerically stiff region. LTE denotes local truncation error, which measures numerical integration accuracy. The Stiffness Index measures the numerical stiff￾ness of the model. (b) Comparison of ten-step forecasting error starting from various timesteps. causing instability fo… 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 4 Pith papers

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

  1. SpecEdit: Training-Free Acceleration for Diffusion based Image Editing via Semantic Locking

    cs.CV 2026-05 unverdicted novelty 7.0

    SpecEdit accelerates diffusion-based image editing up to 10x by using a low-resolution draft to identify edit-relevant tokens via semantic discrepancies for selective high-resolution denoising.

  2. Efficient Video Diffusion Models: Advancements and Challenges

    cs.CV 2026-04 unverdicted novelty 7.0

    A survey that groups efficient video diffusion methods into four paradigms—step distillation, efficient attention, model compression, and cache/trajectory optimization—and outlines open challenges for practical use.

  3. ResilPhase: Plug-and-Play Phase Mapping and Noise-Resilient Macro-Trajectory Extrapolation for Diffusion Acceleration

    cs.AI 2026-06 unverdicted novelty 5.0

    ResilPhase accelerates DiT inference via stable ODE macro-trajectory extrapolation with derivative-free barycentric Lagrange extrapolator and bounded phase mapping, claiming SOTA fidelity at high acceleration ratios o...

  4. Dynamic Video Generation: Shaping Video Generation Across Time and Space

    cs.CV 2026-05 unverdicted novelty 5.0

    DVG dynamically selects content-aware spatio-temporal acceleration strategies for diffusion-based video generation, delivering up to 7x speedup with near-lossless quality on models like HunyuanVideo.

Reference graph

Works this paper leans on

33 extracted references · 6 canonical work pages · cited by 4 Pith papers · 4 internal anchors

  1. [2]

    arXiv preprint arXiv:2303.17604

    Token Merging for Fast Stable Diffusion. arXiv preprint arXiv:2303.17604. Chen, J.; Huang, Z.; Zhang, L.; Wang, J.; Li, J.; Li, X.; Liu, J.; and Chen, T

    Pith/arXiv arXiv

  2. [3]

    arXiv preprint arXiv:2310.00426

    PixArt-alpha: Fast Training of Dif- fusion Transformer for High-Fidelity Text-to-Image Synthe- sis. arXiv preprint arXiv:2310.00426. Chen, J.; Huang, Z.; Zhang, L.; Wang, J.; Li, J.; Li, X.; Liu, J.; and Chen, T

    Pith/arXiv arXiv

  3. [6]

    arXiv preprint arXiv:2305.10924

    Structural Pruning for Diffusion Models. arXiv preprint arXiv:2305.10924. Gu, S.; Lugmayr, A.; Danelljan, M.; Fritsche, M.; Lamour, J.; and Timofte, R

    Pith/arXiv arXiv

  4. [11]

    arXiv preprint arXiv:2311.17982

    VBench: Comprehensive Benchmark Suite for Video Gen- erative Models. arXiv preprint arXiv:2311.17982. Kim, S.; Lee, H.; Cho, W.; Park, M.; and Ro, W. W

    Pith/arXiv arXiv

  5. [12]

    arXiv preprint arXiv:2501.11211

    Ditto: Accelerating Diffusion Model via Temporal Value Similarity. arXiv preprint arXiv:2501.11211. Li, X.; Liu, Y .; Lian, L.; Yang, H.; Dong, Z.; Kang, D.; Zhang, S.; and Keutzer, K

    Pith/arXiv arXiv

  6. [13]

    arXiv preprint arXiv:2302.04304

    Q-Diffusion: Quantizing Diffusion Models. arXiv preprint arXiv:2302.04304. Liu, J.; Zou, C.; Lyu, Y .; Chen, J.; and Zhang, L

    Pith/arXiv arXiv

  7. [14]

    A unified diagrammatic approach to quantum transport in few-level junctions for bosonic and fermionic reservoirs: Application to the quantum Rabi model

    From Reusing to Forecasting: Accelerating Diffusion Models with TaylorSeers. arXiv preprint arXiv:2403.06923. Liu, X.; Gong, C.; and Liu, Q

    work page internal anchor Pith review Pith/arXiv arXiv

  8. [16]

    arXiv preprint arXiv:2206.00927

    DPM-Solver: A Fast ODE Solver for Diffusion Probabilis- tic Model Sampling in Around 10 Steps. arXiv preprint arXiv:2206.00927. Ma, X.; Fang, G.; and Wang, X

    Pith/arXiv arXiv

  9. [17]

    arXiv preprint arXiv:2312.00858

    DeepCache: Accelerating Diffusion Models for Free. arXiv preprint arXiv:2312.00858. Peebles, W.; and Xie, S

    Pith/arXiv arXiv

  10. [18]

    arXiv preprint arXiv:2212.09748

    Scalable Diffusion Models with Transformers. arXiv preprint arXiv:2212.09748. Ronneberger, O.; Fischer, P.; and Brox, T

    Pith/arXiv arXiv

  11. [20]

    arXiv preprint arXiv:2205.11487

    Photorealistic Text-to-Image Diffusion Models with Deep Language Understanding. arXiv preprint arXiv:2205.11487. Selvaraju, P.; Ding, T.; Chen, T.; Zharkov, I.; and Liang, L

    Pith/arXiv arXiv

  12. [21]

    arXiv preprint arXiv:2407.01425

    FORA: Fast-Forward Caching in Diffusion Trans- former Acceleration. arXiv preprint arXiv:2407.01425. Sohl-Dickstein, J.; Weiss, E. A.; Maheswaranathan, N.; and Ganguli, S

    Pith/arXiv arXiv

  13. [22]

    arXiv preprint arXiv:1503.03585

    Deep Unsupervised Learning using Nonequilibrium Thermodynamics. arXiv preprint arXiv:1503.03585. Song, J.; Meng, C.; and Ermon, S

    Pith/arXiv arXiv

  14. [23]

    arXiv preprint arXiv:2010.02502

    Denoising Diffu- sion Implicit Models. arXiv preprint arXiv:2010.02502. Song, Y .; Dhariwal, P.; Chen, M.; and Sutskever, I

    Pith/arXiv arXiv 2010

  15. [24]

    arXiv preprint arXiv:2303.01469

    Consistency Models. arXiv preprint arXiv:2303.01469. Sun, X.; et al

    Pith/arXiv arXiv

  16. [25]

    arXiv preprint arXiv:2411.02265

    Hunyuan-Large: An Open-Source MoE Model with 52 Billion Activated Parameters by Tencent. arXiv preprint arXiv:2411.02265. Wimbauer, F.; Wu, B.; Schoenfeld, E.; Dai, X.; Hou, J.; He, Z.; Sanakoyeu, A.; Zhang, P.; Tsai, S.; Kohler, J.; Rupprecht, C.; Cremers, D.; Vajda, P.; and Wang, J

    Pith/arXiv arXiv

  17. [26]

    arXiv preprint arXiv:2312.03209

    Cache Me if You Can: Accelerating Diffusion Models through Block Caching. arXiv preprint arXiv:2312.03209. Xu, J.; Liu, X.; Wu, Y .; Tong, Y .; Li, Q.; Ding, M.; Tang, J.; and Dong, Y

    Pith/arXiv arXiv

  18. [27]

    Yang, Z.; Jiang, H.; Hong, W.; Teng, J.; Zheng, W.; Dong, Y .; Ding, M.; and Tang, J

    ImageReward: Learning and Evaluat- ing Human Preferences for Text-to-Image Generation.arXiv preprint arXiv:2304.05977. Yang, Z.; Jiang, H.; Hong, W.; Teng, J.; Zheng, W.; Dong, Y .; Ding, M.; and Tang, J

    Pith/arXiv arXiv

  19. [28]

    Inf-DiT: Upsampling Any-Resolution Image with Memory-Efficient Diffusion Transformer

    Inf-DiT: Upsampling Any- Resolution Image with Memory-Efficient Diffusion Trans- former. arXiv preprint arXiv:2405.04312. Yuan, Z.; Zhang, H.; Lu, P.; Ning, X.; Zhang, L.; Zhao, T.; Yan, S.; Dai, G.; and Wang, Y

    work page internal anchor Pith review Pith/arXiv arXiv

  20. [29]

    arXiv preprint arXiv:2406.08552

    DiTFastAttn: Atten- tion Compression for Diffusion Transformer Models. arXiv preprint arXiv:2406.08552. Zhang, E.; Tang, J.; Ning, X.; and Zhang, L

    Pith/arXiv arXiv

  21. [30]

    arXiv preprint arXiv:2408.12588

    Real-Time Video Generation with Pyramid Attention Broadcast. arXiv preprint arXiv:2408.12588. Zheng, K.; Lu, C.; Chen, J.; and Zhu, J

    Pith/arXiv arXiv

  22. [31]

    arXiv preprint arXiv:2310.13268

    DPM-Solver- v3: Improved Diffusion ODE Solver with Empirical Model Statistics. arXiv preprint arXiv:2310.13268. Zhu, H.; Tang, D.; Liu, J.; Lu, M.; Zheng, J.; Peng, J.; Li, D.; Wang, Y .; Jiang, F.; Tian, L.; Tiwari, S.; Sirasao, A.; Yong, J.-H.; Wang, B.; and Barsoum, E

    Pith/arXiv arXiv

  23. [32]

    DiP-GO: A Diffusion Pruner via Few-step Gradient Optimization

    DiP-GO: A Dif- fusion Pruner via Few-step Gradient Optimization. arXiv preprint arXiv:2410.16942. Zou, C.; Liu, X.; Liu, T.; Huang, S.; and Zhang, L

    work page internal anchor Pith review Pith/arXiv arXiv

  24. [33]

    arXiv preprint arXiv:2410.05317

    Ac- celerating Diffusion Transformers with Token-wise Feature Caching. arXiv preprint arXiv:2410.05317

    Pith/arXiv arXiv

  25. [2015]

    arXiv preprint arXiv:1505.04597

    U-Net: Convolutional Networks for Biomedical Image Segmenta- tion. arXiv preprint arXiv:1505.04597. Russakovsky, O.; Deng, J.; Su, H.; Krause, J.; Satheesh, S.; Ma, S.; Huang, Z.; Karpathy, A.; Khosla, A.; Bernstein, M. S.; Berg, A. C.; and Fei-Fei, L

    Pith/arXiv arXiv

  26. [2017]

    In Ad- vances in Neural Information Processing Systems 30 (NIPS 2017)

    GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium. In Ad- vances in Neural Information Processing Systems 30 (NIPS 2017). Ho, J.; Jain, A.; and Abbeel, P

    work page 2017

  27. [2019]

    In 2019 IEEE/CVF International Confer- ence on Computer Vision Workshop (ICCVW), 512–520

    DIV8K: DIVerse 8K Resolution Image Dataset. In 2019 IEEE/CVF International Confer- ence on Computer Vision Workshop (ICCVW), 512–520. Hessel, J.; Holtzman, A.; Forbes, M.; Bras, R. L.; and Choi, Y

    work page 2019

  28. [2020]

    arXiv preprint arXiv:2006.11239

    Denoising Diffusion Probabilistic Models. arXiv preprint arXiv:2006.11239. hpcaitech

    Pith/arXiv arXiv 2006

  29. [2021]

    arXiv preprint arXiv:2104.08718

    CLIPScore: A Reference-free Evaluation Metric for Image Captioning. arXiv preprint arXiv:2104.08718. Heusel, M.; Ramsauer, H.; Unterthiner, T.; Nessler, B.; and Hochreiter, S

    Pith/arXiv arXiv

  30. [2022]

    arXiv preprint arXiv:2209.03003

    Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow. arXiv preprint arXiv:2209.03003. Lu, C.; Zhou, Y .; Bao, F.; Chen, J.; Li, C.; and Zhu, J

    Pith/arXiv arXiv

  31. [2023]

    arXiv preprint arXiv:2311.15127

    Stable Video Diffusion: Scaling Latent Video Diffusion Models to Large Datasets. arXiv preprint arXiv:2311.15127. Bolya, D.; and Hoffman, J

    Pith/arXiv arXiv

  32. [2024]

    Self-enhanced mobility enables vortex pattern formation in living matter

    PixArt-Sigma: Scalable and Efficient Diffusion Transformer for High-Resolution Image Synthe- sis. arXiv preprint arXiv:2403.09068. Cheng, X.; Chen, Z.; and Jia, Z

    work page internal anchor Pith review Pith/arXiv arXiv

  33. [2025]

    arXiv preprint arXiv:2502.00433

    CAT Pruning: Cluster-Aware Token Pruning For Text-to-Image Diffusion Models. arXiv preprint arXiv:2502.00433. Fang, G.; Ma, X.; and Wang, X

    Pith/arXiv arXiv

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.