Pith. sign in

REVIEW 4 major objections 6 minor 43 references

Diffusion models waste work on redundant features; OmniCache reuses them hierarchically and cuts inference latency by up to 35% without 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 ↗

T0 review · grok-4.5

2026-07-30 10:43 UTC pith:RK4UVOGQ

load-bearing objection Solid training-free multi-axis cache system with real engineering value; latency claims hold up better than the fidelity half, which rests on noisy single-run FVD and qualitative motion checks. the 4 major comments →

arxiv 2607.23844 v1 pith:RK4UVOGQ submitted 2026-07-26 cs.CV

OmniCache: Multidimensional Hierarchical Feature Caching For Diffusion Models

Zhaoyuan He , Muhammad Muaz , Lili Qiu This is my paper
classification cs.CV
keywords diffusion modelsfeature cachingvideo generationtoken matchingtraining-free accelerationhierarchical reuseinference efficiency
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.

High-resolution image and video diffusion models are slow because they re-run attention-heavy denoisers many times even when intermediate features barely change. This paper argues that four kinds of redundancy—inside a frame, between frames, in local motion, and across neighboring denoising steps—can be turned into free speedups. OmniCache is a training-free hierarchy that matches similar features, skips their recomputation, and puts the cached activations back in their original positions instead of averaging tokens. Spatial features are reused inside temporal layers and temporal features inside spatial layers, with a separate layer-and-timestep schedule for cross-step reuse. On SD3, SVD-XT, and Latte the method reports latency cuts of up to 35%, 25%, and 28% while keeping visual quality and motion coherence under conservative cache ratios.

Core claim

The authors claim that multidimensional hierarchical feature caching—Token, Frame, Block, and Layered Cache—can remove complementary redundancy in diffusion inference without changing weights. Similarity matching selects what to cache; skipped positions are restored from prior-step activations so token order and spatial-temporal layout stay intact. Structure-aware placement (spatial reuse in temporal layers, temporal reuse in spatial layers) plus a budgeted layer–timestep schedule yields 25–35% wall-clock speedups on representative image and video models while matching baseline fidelity.

What carries the argument

OmniCache: a unified hierarchy that uses cosine-similarity matching to mark cacheable tokens, frames, motion blocks, and layer–timestep pairs, skips their computation, and restores positionally consistent cached activations rather than averaging matched features.

Load-bearing premise

Features chosen by similarity from a previous denoising step stay close enough to the true current activations that putting them back does not hurt quality under the chosen cache ratios.

What would settle it

Run SVD-XT or Latte on videos with fast non-repetitive motion or dense independent object motion at the paper’s default ratios and check whether FVD and motion coherence stay near the no-cache baseline while still delivering the claimed latency cut; a clear quality collapse at the same speedup would refute the claim.

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

If this is right

  • Training-free 25–35% latency cuts become available on existing SD3, SVD-XT, and Latte deployments without fine-tuning.
  • Cache-based restore can replace averaging-based token merging when positional order and temporal consistency matter.
  • Spatial-in-temporal and temporal-in-spatial reuse is a concrete design rule for video denoisers.
  • Layer–timestep budgets can be set from profiled quality-loss-per-latency scores instead of fixed intervals alone.
  • OmniCache can be stacked with step-reduction, distillation, quantization, or faster attention for larger combined speedups if ratios are kept conservative.

Where Pith is reading between the lines

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

  • Prompt- or content-adaptive ratios could widen the safe operating range beyond the fixed conservative defaults the paper recommends.
  • The same match-and-restore pattern may transfer to other iterative generative stacks that alternate spatial and temporal attention.
  • If cross-step drift is the main failure mode, pairing OmniCache’s selection with light forecasting or correction of cached values could raise usable cache ratios.
  • Energy and cost savings scale roughly with the reported latency cuts for large-batch cloud image and video serving.

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 / 6 minor

Summary. The paper proposes OmniCache, a training-free inference-time acceleration framework for image and video diffusion models. It identifies four redundancy sources (intra-frame, inter-frame, motion, and denoising-step) and attacks each with a corresponding module: Token Cache, Frame Cache, Block Cache, and a Layered Cache that greedily schedules layer–timestep cache placements by a quality-loss-per-latency score (w = Q/L) with graph-based refinement. Unlike ToMe-style token merging, matched positions are skipped and restored from previous-step cached activations rather than averaged, which the authors argue preserves positional consistency. Structure-aware S2T/T2S reuse (spatial features cached in temporal layers and vice versa) is ablated against S2S/T2T placement. Experiments on SVD-XT, Latte, and SD3-medium report latency reductions of 25%, 28%, and up to 35% respectively, with FVD/FID/CLIP metrics, ratio ablations, interpolation baselines, and Triton kernel profiling in the appendix.

Significance. If the quality claims hold up under variance-aware evaluation, this is a useful contribution to a crowded but practically important area. Strengths worth naming: (i) it is training-free and demonstrated across three architecturally distinct models (U-Net I2V, DiT T2V, DiT T2I); (ii) the cache-restore-instead-of-average design is well-motivated for position-sensitive video models and is ablated against both interpolation controls and S2S/T2T placement controls; (iii) the systems side is real — patch-local matching overhead reductions and custom Triton gather–scatter kernels with profiling (Appendix A) make the wall-clock claims more credible than FLOP-count proxies; (iv) the ratio-sweep ablations honestly identify conservative operating regimes and the Limitations section is candid. The work does not ship code or machine-checked artifacts, and its evaluation practices (single-run distributional metrics, missing video baselines) are currently below what is needed to certify the headline claim.

major comments (4)
  1. [Table 3, §4.3] The quality-preservation half of the main claim rests on single-run FVD numbers whose spread is larger than the effect being asserted. Frame Cache at 20% gives FVD 489.14 and at 40% gives 495.89 — both better than the no-caching baseline (502.68) — while 60% gives 535.07; the full spread across configurations is ~46 points, versus the 0.44-point gap (503.12 vs 502.68) cited as evidence that OmniCache at 50/40/50 is 'nearly identical' to baseline. Cached configurations should cluster around the baseline, not bracket it by tens of points. This pattern indicates run-to-run/sampling variance comparable to or larger than the claimed quality signal, so 'nearly identical FVD' cannot on its own certify preserved quality. The same fragility affects the '11% additional runtime reduction vs ToMeSD at matched quality' comparison (both endpoints single runs, 503.8 vs 503.12 is within noise). Please r
  2. [§4.3 / Figure 8 (Latte)] The 28% latency-reduction claim for Latte appears only in the abstract, §4.3 text, and the qualitative Figure 8. Unlike SD3 (Table 2) and SVD-XT (Table 3), there is no quantitative table for Latte at all — no FVD, no runtime table, no ratio ablation. One of the three headline numbers is therefore unsupported by the presented measurements. A Latte results table with at least latency and a video quality metric is needed.
  3. [§4.3 (baselines)] The only quantitative baselines are ToMeSD (token merging, a different mechanism family) and DeepCache (Table 4). The 2024–2026 training-free video-caching literature the paper itself surveys — TeaCache, PAB, FasterCache, AdaCache, and the forecasting variants — is never evaluated against. Since the central novelty claim is that coordinating multiple reuse axes outperforms single-axis caching, at least one or two of these direct competitors (e.g., TeaCache or FasterCache on SVD-XT/Latte) should be compared quantitatively; otherwise the claimed advantage over the state of the art is untested.
  4. [§4.3 / Figures 6, 8–10 (motion coherence evidence)] 'Motion coherence' is asserted from Figures 6/8/9/10 only, with no same-initial-noise trajectory comparison for any video model. A per-frame LPIPS/SSIM against the uncached trajectory from identical noise, or a standard temporal-consistency metric, would directly test whether cache restoration preserves motion. This matters because the load-bearing premise — that previous-step cached activations substitute adequately for recomputed ones — is exactly what the Limitations section concedes can fail (fast non-repetitive motion), and Block Cache (the motion-redundancy module) is the first module to degrade in Table 3 (528.24 at 50%). The pixel-wise diagnostics in Table 4 (LPIPS/SSIM/PSNR vs uncached SD3) are a good template; an analogous diagnostic is needed on the video side.
minor comments (6)
  1. [Table 2] All cached configurations report CLIP ≈ 20.1–20.4 versus 18.60 for the uncached SD3-medium baseline. A uniform ~1.7-point CLIP improvement from a quality-neutral caching scheme is surprising and suggests a seed or evaluation-protocol difference; please clarify.
  2. [§4.3 vs Table 2 (headline '35%' for SD3)] Table 2 shows 1.35× speedup at 60% token ratio (≈26% latency reduction); the larger 2.28–2.32× figures include Layered Cache. Please state explicitly which configuration produces the headline 35% (and likewise for the 25%/28% figures) so the abstract numbers are traceable to a table row.
  3. [§3.4 (Layered Cache profiling)] The Layered Cache schedule requires profiling Q_{l,t} (via FID/CLIP/FVD) and L_{l,t} per layer–timestep pair. The cost of this profiling pass, and whether a schedule transfers across models, resolutions, or content distributions, is not discussed; this affects the 'training-free / practical' framing.
  4. [Table 4] FID/CLIP on 2,000 MS-COCO 2017 pairs is high-variance; the caveat about cross-table comparison is appreciated, but confidence intervals would help. Also, the DeepCache rows omit the pixel-wise metrics ('–'), which limits the comparison the table is meant to support.
  5. [Table 1 (H.265 bitrate analogy)] The H.265 bitrate analogy is loose (codec bitrate scaling reflects encoder design and psychovisual targets, not directly feature redundancy in a denoiser). Consider softening the claim or moving the table to an appendix.
  6. [Notation and terminology] The '↑L' column header in Table 2 is undefined; rf/rb/rt are used in §4.2 before being tied to Frame/Block/Token ratios in the table captions; the S2T/T2S acronyms in Figure 9 could be defined at first use in §4.3 rather than only in the contribution list.

Circularity Check

0 steps flagged

No significant circularity: empirical systems paper with measured speedups/quality, not a self-closed derivation.

full rationale

OmniCache is a training-free inference engineering paper. Its load-bearing claims (latency cuts of ~25–35% on SVD-XT/Latte/SD3 with maintained FVD/FID/CLIP) are external measurements against pretrained baselines and held-out prompts/datasets (UCF101, MS-COCO), not quantities forced by rewriting fitted identities as predictions. Token/Frame/Block Cache use similarity matching to skip compute and restore prior-step activations; Layered Cache profiles empirical Q_{l,t}/L_{l,t} and greedily packs a latency budget (Algs. 1–2)—ordinary hyperparameter/scheduling engineering, not a first-principles derivation that reduces to its inputs by construction. Related-work citations (DeepCache, ToMe/ToMeSD, Δ-DiT, etc.) supply baselines and motivation; none is a self-authored uniqueness theorem that forbids alternatives or smuggles the central result. Structure-aware S2T/T2S placement and cache-vs-interpolation ablations are empirical comparisons. No equation equates a claimed prediction to a fitted parameter; no self-definitional loop. Per the analyzer default, this is an honest non-finding.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 2 invented entities

The central speed/quality claim rests on domain assumptions about feature redundancy, hand-chosen operating ratios and schedules, and engineering modules defined by the paper. There is no deep axiomatic physics content; the ledger is mostly hyperparameters and standard diffusion/transformer assumptions.

free parameters (5)
  • Frame/Block/Token cache ratios (rf, rb, rt) = model-specific; e.g. SVD-XT 50/40/50, Latte 30/40/20, SD3 token 40–60%
    Primary quality–speed knobs chosen via ablation; recommended defaults (e.g., rf≤40%, rb≤30%, rt≤40%) and model-specific settings (SVD 50/40/50, Latte 30/40/20) are selected to keep metrics stable.
  • Layered Cache target speedup sr and interval N = N=5 in main Layered Cache sweep; sr varies ~1.6–2.8x
    Budget and reuse interval for greedy layer–timestep selection; N=5 used in SD3 table; directly determines which pairs are cached.
  • Quality/latency scores Q_l,t and L_l,t used in w=Q/L = profile-dependent (not released as a table of raw values)
    Profiled per layer–timestep on MS-COCO prompts; the ranking that builds the cache schedule depends on these measured quantities and chosen metrics (FID/CLIP/FVD).
  • Patch size k for hierarchical local matching = unspecified single default; hierarchy may coarsen after local pass
    Controls similarity-search scope/overhead; paper reports large overhead reductions from patch-local matching but treats k as an implementation choice.
  • Layer inclusion/exclusion masks = e.g. SD3 exclude top/bottom four blocks; SVD avoid last decoder Frame Cache and FFN Frame/Token Cache
    Early/late layers and some FFN blocks excluded because they are quality-sensitive; these hand constraints materially enable the reported quality retention.
axioms (5)
  • domain assumption Standard diffusion denoising formulation and iterative solver evaluations dominate inference cost.
    Section 3.1 background; underpins why skipping intermediate feature compute yields wall-clock gains.
  • domain assumption Natural image/video intermediate features exhibit substantial intra-frame, inter-frame, motion, and cross-step redundancy measurable by cosine similarity.
    Introduction taxonomy and §3.2 modules; without this, cache hits would destroy quality.
  • ad hoc to paper Restoring skipped positions from previous-step cached activations preserves spatial-temporal order better than averaging matched tokens for position-sensitive video DiTs/U-Nets.
    Core design thesis vs ToMeSD in §3.1–3.2 and Figure 2; empirically argued but not proved generally.
  • ad hoc to paper Greedy selection by quality-loss per latency (w=Q/L) plus graph refinement is an adequate cache schedule under a speedup budget.
    Algorithms 1–2 / §3.4; authors call it knapsack-style approximation and note non-optimality in Limitations.
  • domain assumption Cosine similarity on reshaped descriptors/blocks/tokens is a sufficient redundancy signal for cache decisions.
    Used throughout Token/Frame/Block Cache matching; inherited from token-merging literature.
invented entities (2)
  • OmniCache hierarchy (Token Cache + Frame Cache + Block Cache + Layered Cache) no independent evidence
    purpose: Unified multidimensional feature-reuse system for training-free diffusion acceleration.
    Methodological composite rather than a physical entity; defined by the paper’s modules and scheduling policy.
  • Structure-aware S2T/T2S reuse rule no independent evidence
    purpose: Place spatial caches in temporal layers and temporal caches in spatial layers to preserve appearance/motion.
    Design rule tested in Figure 9; specific to this paper’s framing of layer roles.

pith-pipeline@v1.2.0-grok45-kimik3 · 18859 in / 3885 out tokens · 81563 ms · 2026-07-30T10:43:20.807283+00:00 · methodology

0 comments
read the original abstract

High-resolution image and video diffusion models, including SD3, FLUX, and recent video diffusion transformers, have substantially improved generative quality but remain expensive at inference time because they repeatedly evaluate attention-heavy denoisers over many sampling steps. We address this inefficiency by exploiting redundancy in intermediate diffusion features rather than changing model weights or retraining. We identify four complementary redundancy sources in image and video generation: intra-frame, inter-frame, motion, and denoising-step redundancy. Based on this analysis, we propose OmniCache, a unified hierarchical caching framework that performs multidimensional feature reuse through Token Cache, Frame Cache, Block Cache, and Layered Cache. Unlike token-merging baselines that average matched features, OmniCache uses similarity matching to select cacheable features, skips redundant computation, and restores positionally consistent cached activations, preserving feature order and spatial-temporal structure. The resulting framework reuses spatial features in temporal layers and temporal features in spatial layers, while Layered Cache captures cross-step redundancy at the model-layer level. Across SD3, SVD-XT, and Latte, OmniCache reduces inference latency by up to 35%, 25%, and 28%, respectively, while maintaining visual fidelity and motion coherence in a training-free setting.

Figures

Figures reproduced from arXiv: 2607.23844 by Lili Qiu, Muhammad Muaz, Zhaoyuan He.

Figure 1
Figure 1. Figure 1: OmniCache reduces redundancy in a hierarchical cascade across time and space. It caches features across consecutive denoising steps, exploits motion and inter-frame redundancy for non-cached features, and then identifies intra-frame redundancy among spatial tokens. • Intra-frame redundancy: Repeated patterns within a frame create similar spatial tokens. Selecting representative tokens and reusing cached ac… view at source ↗
Figure 2
Figure 2. Figure 2: Limitation of averaging-based token merging in video diffusion. Cache-based restoration preserves positional consistency [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Illustration of OmniCache’s Frame Cache, Block Cache, and Token Cache. Each module uses similarity matching to select cacheable features, skips redundant computation, and restores cached activations at their original positions. denoising steps. This distinction is central: ToMe-style methods merge features by averaging, whereas OmniCache reuses positionally consistent cached features. 3.2 OmniCache We pres… view at source ↗
Figure 4
Figure 4. Figure 4: Illustration of Hierarchical Caching. Similarity search and feature reuse are applied locally within patches before optionally [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: (Left) Algorithm for generating cache schedule proposals. (Right) Refinement process for optimizing cache strategies. into k*k patches, reshape the feature map to (B, T, k, H/k, k, W/k, C), and then permute it to (B*k*k, T, (H/k)*(W/k)*C) for local similarity computation. Patch-local matching reduces Frame Cache overhead by more than 20× on SVD-XT without degrading visual quality or motion. The same hierar… view at source ↗
Figure 6
Figure 6. Figure 6: Performance of OmniCache on SVD-XT. With rf , rb, and rt set to (50%, 30%, 50%), OmniCache is compared against ToMeSD at a comparable 65% token-reduction setting. OmniCache closely matches original SVD-XT quality while improving inference time by 25% [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Visualization examples for SD3-medium with hierarchical Token Cache, Layered Cache, and their combination. Token [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Performance of OmniCache on Latte. With rf , rb, and rt set to (30%, 40%, 20%), OmniCache is compared against ToMeSD at a comparable 25% token-merging setting. OmniCache closely matches original Latte quality while improving inference time by 28%. Method Latency and Speedup Evaluation Metrics SDiT ↑ LDiT (ms) Retrain FID ↓ CLIP ↑ SD3-medium 1.00 138.36 ✗ 31.57 18.60 OmniCache 40% 1.20 114.96 ✗ 31.42 20.26 … view at source ↗
Figure 9
Figure 9. Figure 9: S2T/T2S versus S2S/T2T reuse on SVD-XT with [PITH_FULL_IMAGE:figures/full_fig_p011_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Comparison of Frame Cache, Block Cache, and Token Cache with frame interpolation (FI), block interpolation (BI), and [PITH_FULL_IMAGE:figures/full_fig_p012_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Comparison of our custom Triton kernel implementation against the PyTorch baseline on SD3-medium under different [PITH_FULL_IMAGE:figures/full_fig_p016_11.png] view at source ↗

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

43 extracted references · 24 linked inside Pith

  1. [1]

    Flux: High-fidelity text-to-image and image-to-image generation

    Black Forest Labs . Flux: High-fidelity text-to-image and image-to-image generation. https://bfl.ai, 2025. Accessed: 2025-02-06

    2025

  2. [2]

    Stable video diffusion: Scaling latent video diffusion models to large datasets

    Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, 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

  3. [3]

    Token merging for fast stable diffusion

    Daniel Bolya and Judy Hoffman. Token merging for fast stable diffusion. CVPR Workshop on Efficient Deep Learning for Computer Vision, 2023

    2023

  4. [4]

    Token merging: Your ViT but faster

    Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your ViT but faster. In International Conference on Learning Representations, 2023

    2023

  5. [5]

    Video generation models as world simulators

    Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. 2024. URL https://openai.com/research/video-generation-models-as-world-simulators

    2024

  6. [6]

    Videocrafter2: Overcoming data limitations for high-quality video diffusion models, 2024 a

    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, 2024 a

    2024

  7. [7]

    -dit: A training-free acceleration method tailored for diffusion transformers, 2024 b

    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, 2024 b . URL https://arxiv.org/abs/2406.01125

    Pith/arXiv arXiv 2024

  8. [8]

    Veo: Our first text-to-video model

    DeepMind . Veo: Our first text-to-video model . https://deepmind.google/models/veo/, 2024. Accessed: 2025-06-18

    2024

  9. [9]

    Diffusion models beat gans on image synthesis

    Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34: 0 8780--8794, 2021

    2021

  10. [10]

    Scaling rectified flow transformers for high-resolution image synthesis, 2024

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, Kyle Lacey, Alex Goodwin, Yannik Marek, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis, 2024. URL https://arxiv.org/abs/2403.03206

    Pith/arXiv arXiv 2024

  11. [11]

    Animatediff: Animate your personalized text-to-image diffusion models without specific tuning

    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. International Conference on Learning Representations, 2024

    2024

  12. [12]

    H.264 and H.265 Recommended Bit Rate at General Resolutions , 2024

    Hikvision . H.264 and H.265 Recommended Bit Rate at General Resolutions , 2024. URL https://www.hikvision.com/content/dam/hikvision/ca/faq-document/H.2645-&-H.2645-Recommended-Bit-Rate-at-General-Resolutions.pdf. [Online; accessed 18-May-2024]

    2024

  13. [13]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33: 0 6840--6851, 2020

    2020

  14. [14]

    Ryoo, and Tian Xie

    Kumara Kahatapitiya, Haozhe Liu, Sen He, Ding Liu, Menglin Jia, Chenyang Zhang, Michael S. Ryoo, and Tian Xie. Adacache: Adaptive caching for faster video generation with diffusion transformers, 2024. URL https://arxiv.org/abs/2411.02397

    Pith/arXiv arXiv 2024

  15. [15]

    Elucidating the design space of diffusion-based generative models, 2022

    Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models, 2022. URL https://arxiv.org/abs/2206.00364

    Pith/arXiv arXiv 2022

  16. [16]

    Frequency-aware error-bounded caching for accelerating diffusion transformers, 2026

    Guandong Li. Frequency-aware error-bounded caching for accelerating diffusion transformers, 2026. URL https://arxiv.org/abs/2603.05315

    arXiv 2026

  17. [17]

    Vidtome: Video token merging for zero-shot video editing

    Xirui Li, Chao Ma, Xiaokang Yang, and Ming-Hsuan Yang. Vidtome: Video token merging for zero-shot video editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024

    2024

  18. [18]

    Adacorrection: Adaptive offset cache correction for accurate diffusion transformers, 2026

    Dong Liu, Yanxuan Yu, Ben Lengerich, and Ying Nian Wu. Adacorrection: Adaptive offset cache correction for accurate diffusion transformers, 2026. URL https://arxiv.org/abs/2602.13357

    Pith/arXiv arXiv 2026

  19. [19]

    Timestep embedding tells: It's time to cache for video diffusion model, 2024

    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, 2024. URL https://arxiv.org/abs/2411.19108

    Pith/arXiv arXiv 2024

  20. [20]

    From reusing to forecasting: Accelerating diffusion models with TaylorSeers , 2025

    Jiacheng Liu, Chang Zou, Yuanhuiyi Lyu, Junjie Chen, and Linfeng Zhang. From reusing to forecasting: Accelerating diffusion models with TaylorSeers , 2025. URL https://arxiv.org/abs/2503.06923

    Pith/arXiv arXiv 2025

  21. [21]

    Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps, 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, 2022. URL https://arxiv.org/abs/2206.00927

    Pith/arXiv arXiv 2022

  22. [22]

    Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models

    Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. Machine Intelligence Research, June 2025. ISSN 2731-5398. doi:10.1007/s11633-025-1562-4. URL http://dx.doi.org/10.1007/s11633-025-1562-4

    doi:10.1007/s11633-025-1562-4 2025

  23. [23]

    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, 2024. URL https://arxiv.org/abs/2410.19355

    Pith/arXiv arXiv 2024

  24. [24]

    Latte: Latent diffusion transformer for video generation

    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. arXiv preprint arXiv:2401.03048, 2024 a

    Pith/arXiv arXiv 2024

  25. [25]

    Deepcache: Accelerating diffusion models for free, 2023

    Xinyin Ma, Gongfan Fang, and Xinchao Wang. Deepcache: Accelerating diffusion models for free, 2023. URL https://arxiv.org/abs/2312.00858

    Pith/arXiv arXiv 2023

  26. [26]

    Learning-to-cache: Accelerating diffusion transformer via layer caching

    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: 0 133282--133304, 2024 b

    2024

  27. [27]

    Conditional image-to-video generation with latent flow diffusion models

    Haomiao Ni, Changhao Shi, Kai Li, Sharon X Huang, and Martin Renqiang Min. Conditional image-to-video generation with latent flow diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18444--18455, 2023

    2023

  28. [28]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195--4205, 2023

    2023

  29. [29]

    Progressive distillation for fast sampling of diffusion models

    Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022

    Pith/arXiv arXiv 2022

  30. [30]

    Relational feature caching for accelerating diffusion transformers, 2026

    Byunggwan Son, Jeimin Jeon, Jeongwoo Choi, and Bumsub Ham. Relational feature caching for accelerating diffusion transformers, 2026. URL https://arxiv.org/abs/2602.19506

    arXiv 2026

  31. [31]

    Denoising diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020

    Pith/arXiv arXiv 2010

  32. [32]

    Open-sora: Democratizing efficient video production for all, 2023

    HPC-AI Tech. Open-sora: Democratizing efficient video production for all, 2023. URL https://github.com/hpcaitech/Open-Sora/tree/main. Accessed: 2024-05-21

    2023

  33. [33]

    Sv3d: Novel multi-view synthesis and 3d generation from a single image using latent video diffusion, 2024

    Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. Sv3d: Novel multi-view synthesis and 3d generation from a single image using latent video diffusion, 2024. URL https://arxiv.org/abs/2403.12008

    Pith/arXiv arXiv 2024

  34. [34]

    Microcinema: A divide-and-conquer approach for text-to-video generation

    Yanhui Wang, Jianmin Bao, Wenming Weng, Ruoyu Feng, Dacheng Yin, Tao Yang, Jingxu Zhang, Qi Dai Zhiyuan Zhao, Chunyu Wang, Kai Qiu, et al. Microcinema: A divide-and-conquer approach for text-to-video generation. arXiv preprint arXiv:2311.18829, 2023

    Pith/arXiv arXiv 2023

  35. [35]

    Cache me if you can: Accelerating diffusion models through block caching

    Felix Wimbauer, Bichen Wu, Edgar Schoenfeld, Xiaoliang Dai, Ji Hou, Zijian He, Artsiom Sanakoyeu, Peizhao Zhang, Sam Tsai, Jonas Kohler, et al. Cache me if you can: Accelerating diffusion models through block caching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6211--6220, 2024

    2024

  36. [36]

    Efficient video diffusion models via content-frame motion-latent decomposition

    Sihyun Yu, Weili Nie, De-An Huang, Boyi Li, Jinwoo Shin, and Anima Anandkumar. Efficient video diffusion models via content-frame motion-latent decomposition. ICLR 2024, abs/2403.14148, 2024. URL https://api.semanticscholar.org/CorpusID:268554134

    Pith/arXiv arXiv 2024

  37. [37]

    AB-Cache : Training-free acceleration of diffusion models via Adams-Bashforth cached feature reuse, 2025

    Zichao Yu, Zhen Zou, Guojiang Shao, Chengwei Zhang, Shengze Xu, Jie Huang, Feng Zhao, Xiaodong Cun, and Wenyi Zhang. AB-Cache : Training-free acceleration of diffusion models via Adams-Bashforth cached feature reuse, 2025. URL https://arxiv.org/abs/2504.10540

    Pith/arXiv arXiv 2025

  38. [38]

    Blockdance: Reuse structurally similar spatio-temporal features to accelerate diffusion transformers, 2025

    Hui Zhang, Tingwei Gao, Jie Shao, and Zuxuan Wu. Blockdance: Reuse structurally similar spatio-temporal features to accelerate diffusion transformers, 2025. URL https://arxiv.org/abs/2503.15927

    Pith/arXiv arXiv 2025

  39. [39]

    SoftCap : Soft-budget control for diffusion transformer acceleration, 2026

    Yuhang Zhang, Junxiang Qiu, Huixia Ben, Zhenhua Tang, Shuo Wang, and Yanbin Hao. SoftCap : Soft-budget control for diffusion transformer acceleration, 2026. URL https://arxiv.org/abs/2605.27075

    Pith/arXiv arXiv 2026

  40. [40]

    Real-time video generation with pyramid attention broadcast, 2024

    Xuanlei Zhao, Xiaolong Jin, Kai Wang, and Yang You. Real-time video generation with pyramid attention broadcast, 2024. URL https://arxiv.org/abs/2408.12588

    Pith/arXiv arXiv 2024

  41. [41]

    Less is enough: Training-free video diffusion acceleration via runtime-adaptive caching, 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, 2025. URL https://arxiv.org/abs/2507.02860

    Pith/arXiv arXiv 2025

  42. [42]

    Accelerating diffusion transformers with token-wise feature caching, 2025

    Chang Zou, Xuyang Liu, Ting Liu, Siteng Huang, and Linfeng Zhang. Accelerating diffusion transformers with token-wise feature caching, 2025. URL https://arxiv.org/abs/2410.05317

    Pith/arXiv arXiv 2025

  43. [43]

    DisCa : Accelerating video diffusion transformers with distillation-compatible learnable feature caching, 2026

    Chang Zou, Changlin Li, Yang Li, Patrol Li, Jianbing Wu, Xiao He, Songtao Liu, Zhao Zhong, Kailin Huang, and Linfeng Zhang. DisCa : Accelerating video diffusion transformers with distillation-compatible learnable feature caching, 2026. URL https://arxiv.org/abs/2602.05449

    Pith/arXiv arXiv 2026