Pith. sign in

REVIEW 4 major objections 5 minor 42 references

Video diffusion caching's speed-quality tradeoff is not fundamental; a dual-threshold switch based on the drift signal's rate of change expands the Pareto frontier beyond any fixed threshold.

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

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

T0 review · deepseek-v4-flash

2026-08-02 06:33 UTC pith:ONG7YGW2

load-bearing objection A plausible adaptive-caching wrapper with an overbroad headline claim: the fixed-threshold baselines are too sparse and the same prompts are used to tune and evaluate, so the Pareto expansion is not yet established. the 4 major comments →

arxiv 2607.12358 v2 pith:ONG7YGW2 submitted 2026-07-14 cs.CV

ACID: Adaptive Caching for vIDeo generation

Om Agrawal , Saurabh Agarwal , Aditya Akella This is my paper
classification cs.CV
keywords adaptive cachingvideo diffusion modelsinference accelerationfeature cachingcritical stepsdynamic thresholdtraining-freePareto frontier
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper claims that the apparent tradeoff between visual quality and inference speed in dynamic caching methods for video diffusion is not fundamental—it is an artifact of holding a single caching threshold constant throughout denoising. ACID switches between a low and a high threshold depending on how fast the method's drift signal is changing, applying the low threshold only at critical steps where the signal changes rapidly. This recovers most of the quality of conservative caching while achieving speedups close to aggressive caching, expanding the Pareto frontier in all seven method–model combinations tested. If true, this means every existing fixed-threshold caching method is leaving quality and latency on the table, and the improvement comes for free—no retraining, no change to the base method's core mechanism.

Core claim

On its own terms, the paper establishes that the fixed-threshold design shared by current dynamic feature caching methods is the source of the quality-latency tradeoff, not the caching signal itself. ACID detects critical steps by monitoring the rate of change of the method's existing drift signal and switches between a low caching threshold τ_low (during dynamic, quality-critical phases) and a high threshold τ_high (during stable phases). Concretely, on TeaCache with HunyuanVideo, the adaptive mode reaches 2.16× speedup with PSNR 24.41, compared with 1.57×/24.69 for the fixed low threshold and 3.20×/19.98 for the fixed high threshold. The pattern holds across TeaCache, EasyCache, and DiCach

What carries the argument

The central mechanism is the adaptive dual-threshold controller: maintain two thresholds, τ_high and τ_low; at each denoising step compute the base method's drift signal s_t; if |s_t − s_{t−1}| ≤ δ, classify the phase as stable and apply τ_high for aggressive caching, otherwise apply τ_low. The paper identifies the switch points as aligning with where the second derivative of the drift signal stabilizes, so the signal's rate of change acts as a proxy for quality-criticality. The controller is signal-agnostic—it works with TEMNI for TeaCache, k×Δinput/‖v‖ for EasyCache, and probe-feature differences for DiCache—and requires no retraining or modification of the base caching algorithm.

Load-bearing premise

The load-bearing premise is that the local rate of change of the caching drift signal reliably marks quality-critical steps; if the second-derivative stabilization point does not correspond to actual output-quality importance, then the threshold-switching schedule is arbitrary and the reported Pareto expansion could be an artifact of fitting threshold pairs to the evaluation set.

What would settle it

A concrete experiment would compare ACID's adaptive schedule against a random-switch baseline that changes thresholds at randomly selected timesteps using the same low/high threshold pairs and the same overall caching rate. If the random-switch baseline matches ACID's quality–latency points, the critical-step detector is not doing causal work and the claimed Pareto expansion is not attributable to detecting critical steps. Alternatively, hold the distribution of low-threshold steps fixed but scramble their locations; if quality does not degrade, the timing of the critical-step protection is no

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

If this is right

  • Any dynamic caching method that uses a single fixed threshold throughout denoising is suboptimal; switching thresholds based on the drift signal's rate of change improves the speed–quality Pareto frontier.
  • ACID is a training-free, modular wrapper that can be dropped onto existing dynamic caching methods without altering their core reuse or reconstruction logic.
  • The critical-step structure is consistent across prompts but model-dependent, meaning the adaptive switching rule helps across differing signal shapes (U-shaped versus other patterns) without per-model redesign.
  • On TeaCache/HunyuanVideo, ACID delivers up to 38% additional speedup over the conservative fixed-threshold baseline with negligible quality loss (<0.3 dB PSNR, <0.01 SSIM, <0.01 LPIPS).
  • The benefit appears across three different caching signals, implying the limitation is structural to fixed-threshold design rather than an artifact of any single drift proxy.

Where Pith is reading between the lines

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

  • If the rate-of-change heuristic is a general marker of quality-critical phases, it could also inform other per-step decisions in diffusion sampling—such as where to spend extra compute in ODE solvers, where to apply CFG, or where to skip attention layers—though the paper only tests caching thresholds.
  • The model-dependence of the signal shape suggests that a deployment would need to recalibrate δ and the threshold pair for each new model; the rule that works for HunyuanVideo may not transfer unchanged to CogVideoX.
  • One testable extension is to ablate which steps are protected by the low threshold: if the critical-step locations correspond to perceptually meaningful phases (e.g., layout formation versus texture refinement), the switching schedule could be predicted from the prompt or the latent's statistics rather than from the signal's second derivative.
  • Because the threshold pairs and δ in the paper are selected from the evaluation set, a stronger claim would require a self-tuning rule that derives the switching schedule online from the signal alone without per-dataset hyperparameter search.

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 paper proposes ACID, a training-free wrapper around dynamic caching methods (TeaCache, EasyCache, DiCache) that adaptively switches between a low and a high caching threshold during the denoising process. The central claim is that the quality-latency tradeoff of fixed-threshold caching is not fundamental but an artifact of holding the threshold constant, and that adaptive thresholding expands the Pareto frontier of visual quality versus inference speed beyond any fixed threshold. The method monitors the rate of change of the base caching signal (e.g., TEMNI for TeaCache) and switches to a high threshold in 'stable' phases and a low threshold in 'critical' phases. The paper reports experiments on three video diffusion models and three caching methods, with speedups and fidelity metrics (PSNR/SSIM/LPIPS) against the uncached baseline.

Significance. If the central claim were fully established, the paper would make a useful and practical contribution: a simple, model-agnostic, training-free mechanism that yields strictly better speed-quality tradeoffs than fixed-threshold caching. The appendix contains unusually detailed per-combination tables, which is a strength, and the core idea of modulating the threshold based on the drift signal's dynamics is plausible and easy to implement. However, the load-bearing evidence for 'beyond any fixed threshold' is currently undermined by sparse fixed-threshold sweeps, in-sample hyperparameter selection, and an inconsistency between the algorithmic rule and the text's motivational language. The significance is therefore conditional on additional validation with held-out prompts and denser fixed-threshold baselines.

major comments (4)
  1. [§4.2 and Appendix B, Tables 2–8] The central claim that ACID 'consistently expands the Pareto frontier ... beyond what any fixed threshold achieves' is not supported by the reported fixed-threshold sweeps. For TeaCache/HunyuanVideo (Table 2), only fixed thresholds 0.1, 0.2, 0.3 are reported; the adaptive point (τℓ=0.1, τh=0.3) at 2.16x/24.41 PSNR lies between fixed 0.1 (1.57x/24.69) and fixed 0.2 (2.51x/21.71). Without evaluating a fixed threshold in the interval (0.1, 0.2), one cannot exclude a single fixed τ≈0.15 that achieves comparable or better speed and quality. The same gap appears in most other combinations. More seriously, for EasyCache/HunyuanVideo (Table 5), the advertised adaptive row (0.025,0.045) at 2.39x/30.62 PSNR is dominated on the PSNR-speed axis by fixed τ=0.0375 at 2.42x/30.67 PSNR; the adaptive row has better SSIM/LPIPS, but the paper never defines how the multi-dimensional quality objectives are c
  2. [§4.1 and Appendix B] All hyperparameters (τ_low, τ_high, δ) and the reported 'best Pareto-optimal configuration' are selected on the same 33-prompt VBench set used for evaluation. The paper states that 33 videos per configuration are generated on this set, but gives no held-out prompt split or cross-validation for choosing the adaptive mode. This is in-sample selection: the reported adaptive points may be the best of many tried configurations, and the fixed-threshold baseline is only sparsely sampled. Furthermore, the stability bound δ is never reported for any adaptive row, and the appendix notes that 'Adaptive rows that share the same threshold pair differ in the stability bound δ' without listing its values. This makes the method non-reproducible and the Pareto expansion potentially an artifact of overfitting to the evaluation set. Please provide δ values and validate the chosen configurations on a held-o
  3. [§3.2, §3.3, Algorithm 1] There is a direct inconsistency between the text's motivational signal and the implemented rule. Section 3.2 and the contributions describe the 'second derivative' of the caching signal as the indicator for critical steps (Figure 5), and the contribution bullet states ACID 'switches between low and high thresholds based on the second derivative of each method's metric.' However, Algorithm 1 uses the first difference |s_t − s_{t−1}| ≤ δ to decide stability. These are different operations, and the manuscript never reconciles them. If the intended rule is the first difference, the 'second derivative' language and Figure 5 are misleading; if the intended rule is the second difference, Algorithm 1 is wrong. This is load-bearing because the entire method is defined by that switching rule.
  4. [Abstract and Table 1] The abstract's 'negligible (<0.3 dB PSNR, <0.01 SSIM, <0.01 LPIPS) quality degradation' applies only to the TeaCache/HunyuanVideo combination. For TeaCache/Wan 2.1 (Table 1), the adaptive mode gives PSNR 22.10 versus fixed low threshold 24.93, a 2.83 dB drop; SSIM drops 0.088 and LPIPS rises 0.067. For TeaCache/CogVideoX, the PSNR drops from 37.32 to 32.02. These are not negligible losses, and the abstract's wording can easily be misread as a general property. The claims should be qualified to the specific row where they hold, or the abstract revised to reflect the actual magnitude of quality degradation across configurations.
minor comments (5)
  1. [Table 4] The table lists two 'Adaptive (τℓ=0.1, τh=0.3)' rows with different speedups (1.87x and 1.94x) and no distinguishing label; the appendix says they differ in δ, but δ is not shown. This is confusing and should be made explicit by reporting δ in the table or in a caption.
  2. [§3.3] The phrase 'rate of change' is used loosely. Algorithm 1 uses |s_t − s_{t−1}|, which is a first difference, while the text speaks of 'second derivative.' Please define the signal precisely and use consistent terminology throughout.
  3. [§3.2] The claim that 'the optimal transition points coincide with the step at which the second derivative stabilizes' is supported only by visual inspection of Figure 5. A quantitative definition of 'stabilizes' (e.g., threshold on the second-difference magnitude) and a test of alignment would strengthen the motivation.
  4. [§4.1] All experiments use a single fixed seed. Since video generation is stochastic and the prompt set has only 33 prompts, reporting variance across seeds or at least multiple seeds for the headline configurations would increase confidence in the Pareto comparisons.
  5. [References] The paper cites AdaCache [23] but does not compare to it or discuss how ACID relates to existing adaptive-caching methods beyond noting it is a wrapper. A brief comparison or discussion would position the contribution better.

Circularity Check

1 steps flagged

Headline Pareto expansion is partly an in-sample selection over the evaluation set; no derivation-level circularity or self-citation chain.

specific steps
  1. fitted input called prediction [Section 4.1 / Table 1 caption / Appendix B]
    "For each method and model, rows show: baseline (no caching), fixed low threshold, fixed high threshold, and our adaptive method (best Pareto-optimal configuration). ... For each method–model combination, we generated 33 videos per configuration (one per prompt) across all evaluated threshold settings, using a fixed seed throughout."

    The adaptive hyperparameters (tau_low, tau_high, delta) are swept over many configurations on the same 33-prompt VBench set that is later used for the headline comparison. Table 1 then reports the 'best Pareto-optimal configuration' from that sweep as ACID's result. The claimed 'expansion of the Pareto frontier' is therefore the selection objective evaluated on the same data, not the performance of a fixed, pre-specified ACID configuration. The advertised adaptive point is an in-sample maximum over the grid; it is not an independent prediction that a chosen ACID will generalize to new prompts. Because delta is never reported and no held-out prompt split is used, the reported point is a fitted statistic rather than an estimate of a specified method.

full rationale

The only candidate for circularity is the selection of the reported adaptive configuration. Section 4.1 states that all configurations are evaluated on the same 33 prompts, and Table 1's caption calls the adaptive row the 'best Pareto-optimal configuration.' Since tau_low, tau_high, and delta are swept on that set and the best is then used as evidence for the abstract's claim that ACID 'consistently expands the Pareto frontier,' the headline point is an in-sample maximum rather than the performance of a fixed pre-specified method. This is a fitted-input-called-prediction pattern, though it is not a full derivation collapse: the appendix shows several adaptive rows are not Pareto-improving (e.g., EasyCache/HunyuanVideo: fixed tau=0.0375 at 2.42x/30.67 PSNR beats the advertised adaptive 2.39x/30.62), so the selection does not force the advertised result. There are no self-citations or imported uniqueness theorems, and Algorithm 1's operation is fully specified in terms of the base methods' signals. The unreported delta, the first-difference/second-derivative mismatch, and the sparse fixed-threshold sweep are correctness/reproducibility concerns, not circularity.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 1 invented entities

The central claim rests on three tuned quantities (tau_low, tau_high, delta), a qualitative link between signal rate-of-change and quality-criticality, and the assumption that the base caching signals faithfully track output change. The invented 'critical steps' entity has no independent handle outside the paper's own evaluation set. These are the real costs the reader pays for the claimed Pareto expansion.

free parameters (4)
  • tau_low = Per method/model, e.g., 0.1 for TeaCache+HunyuanVideo, 0.05 for TeaCache+Wan 2.1, 0.02 for EasyCache+Wan 2.1
    Low caching threshold used during critical phases; grid-searched on the evaluation set and reported per combination.
  • tau_high = Per method/model, e.g., 0.3 for TeaCache+HunyuanVideo, 0.2 for TeaCache+Wan 2.1, 0.225 for DiCache+Wan 2.1
    High caching threshold used during stable phases; grid-searched on the evaluation set.
  • delta = Not reported in the paper
    Stability bound in Algorithm 1 that decides when to switch between tau_low and tau_high. It is load-bearing and method/model-dependent, but no values are given.
  • reported adaptive mode selection = Best Pareto-optimal row from the scanned threshold/delta grid per combination
    The headline adaptive number is the best of many configurations in Appendix B, chosen after seeing the evaluation results; this selection is itself a tuning choice.
axioms (5)
  • standard math Standard Gaussian diffusion forward/reverse process (Eq. 1-2)
    Background formalism used to set up the denoising problem; not proved in the paper.
  • ad hoc to paper The caching signal's rate of change identifies quality-critical denoising steps
    Central heuristic in Section 3.2-3.3; supported only by visual alignment in Figures 4-5, with no quantitative criterion.
  • ad hoc to paper A single delta on |s_t - s_{t-1}| cleanly separates stable from critical phases across three models
    Algorithm 1 depends on this threshold, but delta values are not reported and transferability is not tested.
  • domain assumption The base methods' drift signals (TEMNI, k*Delta_input/||v||, probe features) correlate with true model output change
    Inherited from TeaCache/EasyCache/DiCache; ACID's switching signal is only as valid as these base signals.
  • domain assumption 33 VBench prompts with a fixed seed are representative for quality-latency ranking
    Evaluation design in Section 4.1; there is no variance estimate and no held-out prompt set for hyperparameter selection.
invented entities (1)
  • critical steps no independent evidence
    purpose: Timesteps where the caching signal changes rapidly; ACID applies the low threshold at these steps to preserve quality.
    The concept is defined and validated on the same signals and models used for evaluation. The paper provides no independent falsifiable measurement that these steps are quality-critical, e.g., via a separate ablation that measures per-step error.

pith-pipeline@v1.3.0-alltime-deepseek · 13856 in / 15664 out tokens · 166606 ms · 2026-08-02T06:33:13.027273+00:00 · methodology

0 comments
read the original abstract

Video diffusion models produce high-quality generations but remain slow at inference due to their sequential denoising procedure. Caching-based acceleration methods address this by reusing intermediate model outputs: leading dynamic approaches such as TeaCache, EasyCache, and DiCache accumulate a drift signal and skip expensive model evaluations when accumulated drift stays below a fixed threshold $\tau$. This threshold controls an apparent tradeoff - raising it yields faster generation at the cost of visual quality, while lowering it preserves quality but sacrifices speed. We show this tradeoff is not fundamental; it is an artifact of holding $\tau$ constant throughout denoising. We identify the existence of critical steps - timesteps where the drift signal changes rapidly - and show that applying a low threshold selectively at these steps while caching aggressively elsewhere recovers most of the quality of conservative caching at substantially higher inference speeds. Building on this insight, we propose ACID, a lightweight, training-free wrapper that monitors the rate of change of each method's existing drift signal to dynamically switch between a low and a high threshold. ACID is signal-agnostic and modular: it requires no retraining and plugs directly into existing dynamic caching methods without modifying their core mechanisms. Evaluated across three caching methods (TeaCache, EasyCache, DiCache) and three open-source video diffusion models (HunyuanVideo, Wan 2.1, CogVideoX), ACID consistently expands the Pareto frontier of visual quality versus inference speed beyond what any fixed threshold achieves. In particular, on TeaCache and HunyuanVideo, ACID achieves up to 2.16x speedup over the no-caching baseline, and up to 38% additional speedup over the conservative fixed-threshold baseline with negligible (<0.3 dB PSNR, <0.01 SSIM, <0.01 LPIPS) quality degradation.

Figures

Figures reproduced from arXiv: 2607.12358 by Aditya Akella, Om Agrawal, Saurabh Agarwal.

Figure 1
Figure 1. Figure 1: DiCache on Wan 2.1. Our adaptive thresholding expands the Pareto frontier beyond any [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Comparison of caching metric consistency across diverse prompts. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Analysis of structural patterns in caching signals across different models. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: ∆TEMNI signal across all three models throughout the denoising process. Vertical dashed lines mark the optimal threshold-transition points identified in the multi-threshold experiment [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Second derivative of the ∆TEMNI signal across all three models, with the same threshold￾transition markers. The transition points align with where the second derivative stabilizes, motivating its use as a signal for adaptive threshold switching. the rate of change of the underlying signal of the dynamic caching method (∆TEMNI for TeaCache, k × ∆input/∥vt−1∥ for EasyCache, and ∆probe features for DiCache). … view at source ↗
Figure 6
Figure 6. Figure 6: TeaCache Pareto frontiers across all three models. Adaptive modes (colored markers) [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: EasyCache Pareto frontiers. Adaptive modes expand the frontier beyond any fixed-threshold [PITH_FULL_IMAGE:figures/full_fig_p013_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: DiCache Pareto frontiers. Adaptive modes expand the frontier beyond any fixed-threshold [PITH_FULL_IMAGE:figures/full_fig_p013_8.png] view at source ↗

Review history (2 revisions) →

discussion (0)

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

Reference graph

Works this paper leans on

42 extracted references · 20 linked inside Pith

  1. [1]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 10684–10695, 2022

    2022

  2. [2]

    Sdxl: Improving latent diffusion models for high-resolution image synthesis.arXiv preprint arXiv:2307.01952, 2023

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis.arXiv preprint arXiv:2307.01952, 2023

    Pith/arXiv arXiv 2023

  3. [3]

    Animatediff: Animate your personalized text-to-image diffusion models without specific tuning.arXiv preprint arXiv:2307.04725, 2023

    Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning.arXiv preprint arXiv:2307.04725, 2023

    Pith/arXiv arXiv 2023

  4. [4]

    Videocrafter2: Overcoming data limitations for high-quality video diffusion models

    Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. Videocrafter2: Overcoming data limitations for high-quality video diffusion models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 7310–7320, 2024

    2024

  5. [5]

    Stable video diffusion: Scaling latent video diffusion models to large datasets.arXiv preprint arXiv:2311.15127, 2023

    Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram V oleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets.arXiv preprint arXiv:2311.15127, 2023

    Pith/arXiv arXiv 2023

  6. [6]

    Scaling rectified flow transformers for high-resolution image synthesis

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. InForty-first International Conference on Machine Learning, 2024

    2024

  7. [7]

    Flux.https://github.com/black-forest-labs/flux, 2024

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

    2024

  8. [8]

    Pixart- α: Fast training of diffusion transformer for photorealistic text-to-image synthesis.arXiv preprint arXiv:2310.00426, 2023

    Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart- α: Fast training of diffusion transformer for photorealistic text-to-image synthesis.arXiv preprint arXiv:2310.00426, 2023

    Pith/arXiv arXiv 2023

  9. [9]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers. InProceedings of the IEEE/CVF International Conference on Computer Vision, pp. 4195–4205, 2023. 9

    2023

  10. [10]

    Flashattention: Fast and memory- efficient exact attention with io-awareness.Advances in Neural Information Processing Systems, 35:16344– 16359, 2022

    Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory- efficient exact attention with io-awareness.Advances in Neural Information Processing Systems, 35:16344– 16359, 2022

    2022

  11. [11]

    Svdquant: Absorbing outliers by low-rank components for 4-bit diffusion models.arXiv preprint arXiv:2411.05007, 2024

    Muyang Li, Yujun Lin, Zhekai Zhang, Tianle Cai, Xiuyu Li, Junxian Guo, Enze Xie, Chenlin Meng, Jun-Yan Zhu, and Song Han. Svdquant: Absorbing outliers by low-rank components for 4-bit diffusion models.arXiv preprint arXiv:2411.05007, 2024

    arXiv 2024

  12. [12]

    Sparse videogen: Accelerating video diffusion transformers with spatial-temporal sparsity.arXiv preprint arXiv:2502.01776, 2025

    Haocheng Xi, Shuo Yang, Yilong Zhao, Chenfeng Xu, Muyang Li, Xiuyu Li, Yujun Lin, Han Cai, Jintao Zhang, Dacheng Li, et al. Sparse videogen: Accelerating video diffusion transformers with spatial-temporal sparsity.arXiv preprint arXiv:2502.01776, 2025

    Pith/arXiv arXiv 2025

  13. [13]

    Grouping first, attending smartly: Training-free acceleration for diffusion transformers.arXiv preprint arXiv:2505.14687, 2025

    Sucheng Ren, Qihang Yu, Ju He, Alan Yuille, and Liang-Chieh Chen. Grouping first, attending smartly: Training-free acceleration for diffusion transformers.arXiv preprint arXiv:2505.14687, 2025

    Pith/arXiv arXiv 2025

  14. [14]

    On distillation of guided diffusion models

    Chenlin Meng, Robin Rombach, Ruiqi Gao, Diederik Kingma, Stefano Ermon, Jonathan Ho, and Tim Salimans. On distillation of guided diffusion models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 14297–14306, 2023

    2023

  15. [15]

    Adversarial diffusion distillation

    Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. InEuropean Conference on Computer Vision, pp. 87–103. Springer, 2024

    2024

  16. [16]

    Videolcm: Video latent consistency model.arXiv preprint arXiv:2312.09109, 2023

    Xiang Wang, Shiwei Zhang, Han Zhang, Yu Liu, Yingya Zhang, Changxin Gao, and Nong Sang. Videolcm: Video latent consistency model.arXiv preprint arXiv:2312.09109, 2023

    Pith/arXiv arXiv 2023

  17. [17]

    Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps.Advances in Neural Information Processing Systems, 35:5775–5787, 2022

    Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps.Advances in Neural Information Processing Systems, 35:5775–5787, 2022

    2022

  18. [18]

    Deepcache: Accelerating diffusion models for free

    Xinyin Ma, Gongfan Fang, and Xinchao Wang. Deepcache: Accelerating diffusion models for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 15762–15772, 2024

    2024

  19. [19]

    Faster diffusion: Rethinking the role of unet encoder in diffusion models.CoRR, 2023

    Senmao Li, Taihang Hu, Fahad Shahbaz Khan, Linxuan Li, Shiqi Yang, Yaxing Wang, Ming-Ming Cheng, and Jian Yang. Faster diffusion: Rethinking the role of unet encoder in diffusion models.CoRR, 2023

    2023

  20. [20]

    Fora: Fast-forward caching in diffusion transformer acceleration.arXiv preprint arXiv:2407.01425, 2024

    Pratheba Selvaraju, Tianyu Ding, Tianyi Chen, Ilya Zharkov, and Luming Liang. Fora: Fast-forward caching in diffusion transformer acceleration.arXiv preprint arXiv:2407.01425, 2024

    Pith/arXiv arXiv 2024

  21. [21]

    ∆-dit: A training-free acceleration method tailored for diffusion transformers.arXiv preprint arXiv:2406.01125, 2024

    Pengtao Chen, Mingzhu Shen, Peng Ye, Jianjian Cao, Chongjun Tu, Christos-Savvas Bouganis, Yiren Zhao, and Tao Chen. ∆-dit: A training-free acceleration method tailored for diffusion transformers.arXiv preprint arXiv:2406.01125, 2024

    Pith/arXiv arXiv 2024

  22. [22]

    Real-time video generation with pyramid attention broadcast.arXiv preprint arXiv:2408.12588, 2024

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

    Pith/arXiv arXiv 2024

  23. [23]

    Ryoo, and Tian Xie

    Kumara Kahatapitiya, Haozhe Liu, Sen He, Ding Liu, Menglin Jia, Chenyang Zhang, Michael S. Ryoo, and Tian Xie. Adaptive caching for faster video generation with diffusion transformers.arXiv preprint arXiv:2411.02397, 2024

    Pith/arXiv arXiv 2024

  24. [24]

    Zhengyao Lv, Chenyang Si, Junhao Song, Zhenyu Yang, Yu Qiao, Ziwei Liu, and Kwan-Yee K. Wong. Fastercache: Training-free video diffusion model acceleration with high quality.arXiv preprint arXiv:2410.19355, 2024

    Pith/arXiv arXiv 2024

  25. [25]

    Classifier-free diffusion guidance.arXiv preprint arXiv:2207.12598, 2022

    Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance.arXiv preprint arXiv:2207.12598, 2022

    Pith/arXiv arXiv 2022

  26. [26]

    Timestep embedding tells: It’s time to cache for video diffusion model

    Feng Liu, Shiwei Zhang, Xiaofeng Wang, Yujie Wei, Haonan Qiu, Yuzhong Zhao, Yingya Zhang, Qixiang Ye, and Fang Wan. Timestep embedding tells: It’s time to cache for video diffusion model. InProceedings of the Computer Vision and Pattern Recognition Conference, pp. 7353–7363, 2025

    2025

  27. [27]

    Less is enough: Training-free video diffusion acceleration via runtime-adaptive caching.arXiv preprint arXiv:2507.02860, 2025

    Xin Zhou, Dingkang Liang, Kaijin Chen, Tianrui Feng, Xiwu Chen, Hongkai Lin, Yikang Ding, Feiyang Tan, Hengshuang Zhao, and Xiang Bai. Less is enough: Training-free video diffusion acceleration via runtime-adaptive caching.arXiv preprint arXiv:2507.02860, 2025

    Pith/arXiv arXiv 2025

  28. [28]

    J. Bu, P. Ling, Y . Zhou, Y . Wang, Y . Zang, D. Lin, and J. Wang. Dicache: Let diffusion model determine its own cache.arXiv preprint arXiv:2508.17356, 2025. 10

    arXiv 2025

  29. [29]

    Open-sora: Democratizing efficient video production for all.arXiv preprint arXiv:2412.20404, 2024

    Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all.arXiv preprint arXiv:2412.20404, 2024

    Pith/arXiv arXiv 2024

  30. [30]

    Latte: Latent diffusion transformer for video generation.Transactions on Machine Learning Research, 2024

    Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Ziwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation.Transactions on Machine Learning Research, 2024

    2024

  31. [31]

    Video generation models as world simulators.OpenAI Blog, 1:8, 2024

    Tim Brooks et al. Video generation models as world simulators.OpenAI Blog, 1:8, 2024

    2024

  32. [32]

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

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

    Pith/arXiv arXiv 2025

  33. [33]

    Cogvideo: Large-scale pretraining for text-to-video generation via transformers

    Wenyi Hong, Ming Ding, Wendi Zheng, Xinghao Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. InProceedings of the International Conference on Learning Representations, 2023

    2023

  34. [34]

    Cogvideox: Text-to-video diffusion models with an expert transformer.arXiv preprint arXiv:2408.06072, 2024

    Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer.arXiv preprint arXiv:2408.06072, 2024

    Pith/arXiv arXiv 2024

  35. [35]

    Vbench: Comprehensive benchmark suite for video generative models

    Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 21807–21818, 2024

    2024

  36. [36]

    Vbench-2.0: Advancing video generation benchmark suite for intrinsic faithfulness.arXiv preprint arXiv:2503.21755, 2025

    Dian Zheng, Ziqi Huang, Hongbo Liu, Kai Zou, Yinan He, Fan Zhang, Lulu Gu, Yuanhan Zhang, Jingwen He, Wei-Shi Zheng, Yu Qiao, and Ziwei Liu. Vbench-2.0: Advancing video generation benchmark suite for intrinsic faithfulness.arXiv preprint arXiv:2503.21755, 2025

    Pith/arXiv arXiv 2025

  37. [37]

    Hunyuanvideo: A systematic framework for large video generative models.arXiv preprint arXiv:2412.03603, 2024

    Weijie Kong et al. Hunyuanvideo: A systematic framework for large video generative models.arXiv preprint arXiv:2412.03603, 2024

    Pith/arXiv arXiv 2024

  38. [38]

    Snapfusion: Text-to-image diffusion model on mobile devices within two seconds.Advances in Neural Information Processing Systems, 36, 2024

    Yanyu Li, Huan Wang, Qing Jin, Ju Hu, Pavlo Chemerys, Yun Fu, Yanzhi Wang, Sergey Tulyakov, and Jian Ren. Snapfusion: Text-to-image diffusion model on mobile devices within two seconds.Advances in Neural Information Processing Systems, 36, 2024

    2024

  39. [39]

    Q-dit: Accurate post-training quantization for diffusion transformers

    Lei Chen, Yuan Meng, Chen Tang, Xinzhu Ma, Jingyan Jiang, Xin Wang, Zhi Wang, and Wenwu Zhu. Q-dit: Accurate post-training quantization for diffusion transformers. InProceedings of the Computer Vision and Pattern Recognition Conference, pp. 28306–28315, 2025

    2025

  40. [40]

    Learning-to-cache: Accelerating diffusion transformer via layer caching.Advances in Neural Information Processing Systems, 37:133282–133304, 2024

    Xinyin Ma, Gongfan Fang, Michael Bi Mi, and Xinchao Wang. Learning-to-cache: Accelerating diffusion transformer via layer caching.Advances in Neural Information Processing Systems, 37:133282–133304, 2024

    2024

  41. [41]

    Spargeattn: Accurate sparse attention accelerating any model inference.arXiv preprint arXiv:2502.18137, 2025

    Jintao Zhang, Chendong Xiang, Haofeng Huang, Jia Wei, Haocheng Xi, Jun Zhu, and Jianfei Chen. Spargeattn: Accurate sparse attention accelerating any model inference.arXiv preprint arXiv:2502.18137, 2025

    arXiv 2025

  42. [42]

    From reusing to forecasting: Accelerating diffusion models with taylorseers.arXiv preprint arXiv:2503.06923, 2025

    Jiacheng Liu, Chang Zou, Yuanhuiyi Lyu, Junjie Chen, and Linfeng Zhang. From reusing to forecasting: Accelerating diffusion models with taylorseers.arXiv preprint arXiv:2503.06923, 2025. A Pareto Frontier Plots Figures 6–8 show Pareto frontier plots for all evaluated method–model combinations. Each plot compares fixed-threshold configurations across a ran...

    Pith/arXiv arXiv 2025