Pith. sign in

REVIEW 2 major objections 6 minor 48 references

GVCCTurbo: Rate-Compute Quality Scheduling for Codebook Driven Generative Compression

T0 review · 2 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read A scheduler that separates bitrate from compute cuts generative-codec decoding time by about 44% at near-equal quality.

desk verdict A clean scheduler for codebook-driven compression with solid NFE savings, but the quality-neutrality claim is mildly undercut by calibration overlap with evaluation content; worth reviewing with a request for disjoint calibration. read the letter →

arxiv 2608.03517 v1 pith:ETDXGAQS submitted 2026-08-04 cs.CV

Ziyue Zeng , Dingjie Peng , Xun Su , Hiroshi Watanabe This is my paper
classification cs.CV
keywords codebook-drivengenerativecompressionrate-computeschedulingendpointcachingpriorrefreshthinningrectifiedflowvideodiffusionimagezero-shotcodecBPP-drivensamplingschedule
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

At ultra-low bitrates, codebook-driven generative compression steers a pretrained generator with compact codebook indices. Until now, each finite-rate correction required a fresh, expensive evaluation of the generator, so shortening the sampler to save compute also deleted correction slots that carry target information. This paper argues that bitrate and compute are separable resources and shows a way to decouple them: after a one-time calibration of two constants per protocol, the scheduler maps a target codebook-payload bitrate to a trajectory length and a refresh period. In a controlled 720p video study, prior evaluations drop from 20 to 9, cutting decoding time by roughly 44% at a small perceptual cost on high-motion content, and the same interface carries over to diffusion-based image compression. The reader should care because the result makes bit-per-pixel a schedule input rather than a fixed consequence of sampler length, opening a zero-training path to faster generative codecs.

What carries the argument

Endpoint-cached prior refresh with a probe-free (T, p, M) calculator. The cache stores the clean-endpoint estimate from the last refresh step, and at skipped steps the codec recomputes the local drift from the current state, so the correction grid stays dense while prior evaluations are thinned. The endpoint residual is shown to be the mis-attributed noise in the stochastic slot (up to a path-specific scalar), which is what lets a fixed-rate codebook projection of the cached residual steer the trajectory. The calculator sets N_code from the target codebook-payload BPP, then T and the refresh period p from the calibrated skip-gap ratio, making BPP the input that drives both trajectory length

What would settle it

Measure endpoint staleness directly at skipped steps: compute the fresh prior evaluation that the schedule omits, form the coded residual with both the cached and the fresh endpoint, and check whether the Top-M atom selections differ or the angular misalignment between residuals exceeds the selection threshold within the calibrated skip gap. If a substantial fraction of steps within the calibrated gap show such divergence, or a held-out sequence with rapid motion collapses in LPIPS at the calibrated gap, the scheduler's quality-maintenance claim would be falsified.

Watch

Extended reading notes

Core claim

GVCCTurbo's central claim is that a codebook-driven generative codec can keep its dense grid of finite-rate corrections while thinning the expensive prior evaluations, provided the two are scheduled through a shared, precommitted interface. The method caches the clean-endpoint prediction from the most recent refresh step and reuses it for several correction steps, recomputing the local velocity from the current state without a new prior evaluation. Because an endpoint error is shown to be identically a noise-attribution error, projecting the cached-endpoint residual onto the shared codebook still steers the trajectory. A probe-free calculator then allocates the number of correction steps fro

Load-bearing premise

The load-bearing premise is that reusing a cached endpoint estimate across a skipped window keeps endpoint staleness small enough that the discrete Top-M codebook selection still picks useful innovation atoms; the paper's own analysis shows this can fail discontinuously once staleness becomes comparable to the residual, and it is guaranteed only by the protocol-calibrated skip-gap ratio, not by a derivation.

Editorial extensions

If this is right

  • Decoding time falls ~44% when prior evaluations drop from 20 to 9, at a small shared LPIPS cost on high-motion content; this saving is shared across the whole schedule family, including the uniform-thinning boundary point.
  • Codebook-payload BPP becomes a schedule input: a target bitrate directly determines trajectory length, refresh period, and atom count, so rate and compute can be traded at run time without retraining or online probes.
  • Uniform refresh thinning is a boundary point of the (T, p, M) family; the BPP-aware interior point uses 2.9% fewer codebook-payload bits with consistently higher PSNR at comparable LPIPS, so the allocated point need not dominate every boundary point.
  • The same endpoint-prediction and finite-rate steering interface covers rectified-flow video and diffusion image compression, with the atom-rate model swapped per backbone.
  • Decoupling the correction count from the prior-evaluation count suggests that a distilled few-step prior could supply a dense correction grid within its native evaluation budget; the paper reports this as a preliminary observation, not an established gain.

Reading between the lines

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

  • The endpoint/noise identity suggests that any codec steering a latent trajectory with directional innovations could reuse cached endpoints, including latent-transform generative codecs that currently couple refinement steps to generator calls; a testable extension is to apply the same (T, p, M) schedule there and measure whether the quality-versus-gap cliff follows the same protocol-calibrated sha
  • The skip-gap cliff being protocol-calibrated rather than derived implies the scheduler's guarantee is tied to backbone step-schedule alignment; for priors with steeper or more non-smooth trajectories, the calibrated gap may need to shrink, so the ~44% saving may not transfer to all backbones without recalibration.
  • The paper's BPP budget excludes side information and container headers; an external total-rate allocator that knows content-dependent side rate could set the codebook budget adaptively, turning the scheduler into a component of a rate-control loop rather than a fixed mapping.
  • The paper notes N_FE tracks wall-clock only when prior evaluations dominate; in small-latent settings per-step correction cost is non-negligible, so a per-step-aware compute model is needed to compare schedules, a direction the paper flags but does not develop.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 6 minor

Summary. The paper proposes GVCCTurbo, a training-free scheduler for codebook-driven generative compression. It separates the number of expensive prior evaluations (N_FE) from the number of finite-rate codebook corrections (N_code) by caching the predicted clean endpoint over a fixed refresh period and reusing it for several correction steps. After calibrating an atom-count operating point M_perc and a skip-gap ratio tau_gap once per protocol, a probe-free calculator maps a target codebook-payload BPP to a trajectory length T and refresh period p. The interface is demonstrated on a GVCC-style rectified-flow video backbone and a Turbo-DDCM-style diffusion image codec. In a controlled 720p Wan-GVCC study, pure-skip and the BPP-aware schedule both reduce N_FE from 20 to 9 with a measured ~44% decoder-time reduction; the BPP-aware point uses 2.9% fewer codebook-payload bits at comparable quality. Native 1080p rate-distortion curves position the full codec in the ultra-low-bitrate regime.

Significance. If the empirical claims hold, the paper introduces a clean and useful abstraction: BPP as a schedule input rather than a consequence of sampler length, with no retraining and no change to the pretrained prior. The rate-allocation formulas are straightforward arithmetic from Eqs. (11)-(13), and the paper is careful to label tau_gap as a calibrated stability bound rather than a derived guarantee. The paper also explicitly limits the N_FE-to-wall-clock mapping to regimes where prior evaluations dominate, and marks the PiD and MR appendices as diagnostics; Appendix B.4 appropriately calls the staleness argument local intuition. These are strengths. The main risk is empirical, not derivational: the two calibrated constants are selected with sweeps that overlap the evaluation benchmarks, so the controlled 20-to-9 'small shared LPIPS cost' may be in-sample. The paper's own five-clip held-out check is suggestive but too small to remove that risk.

major comments (2)
  1. [Tables 2-3 and Supplementary D.2] The controlled Table 1 comparison is not independent of the calibration that produces the schedules. The tau_gap calibration (Table 3) uses a two-sequence stress pair, and the multi-GOP variant in Table 5(b) explicitly adds HoneyBee and videoSRC07; HoneyBee is a UVG sequence and videoSRC07 is an MCL-JCV sequence, both of which appear in the Table 1 evaluation. The M_perc sweep in Table 2 uses the same diagnostic protocol. Since M_perc=48 and tau_gap=0.15 are selected at least partly on benchmark content, the small LPIPS deltas reported for the 20-to-9 schedules in Table 1 may reflect in-sample selection rather than a general property of the scheduler. The held-out check in D.2 (five clips, one GOP each) is a constructive first step but does not establish that the exact Table 1 deltas are selection-independent. This is load-bearing because the headline is not the 44% speedup (which follow
  2. [Calibration protocol specification] The sequences and GOPs used for Table 2's M-sweep and Table 3's stress probe are not disclosed in the main text; Supplement D.2 gives only partial information and, for the multi-GOP sweep, explicitly includes HoneyBee and videoSRC07. Without a full list of calibration material, a reader cannot check whether the constants are independent of the evaluation sets. This is a missing-support issue for the central empirical claim, not a style preference. Please provide the complete sequence/GOP inventory for all calibration sweeps and, if any benchmark material is used, justify why the subsequent controlled evaluation is still valid.
minor comments (6)
  1. [Abstract / Table 1] The phrase 'small shared LPIPS cost on high-motion content' is stronger than Table 1 supports: the UVG delta is -0.001, MCL-JCV is +0.002, and HEVC-B is +0.014. 'Shared' should be qualified as 'on average, with the largest increase on HEVC-B.'
  2. [Eq. (10) and Table 1] For T=20 and tau_gap=0.15, Eq. (10) gives p*=4, yet the pure-skip row uses p=3. Explain that pure-skip is not the calculator's maximal period but a boundary point chosen to keep N_FE=9; otherwise the relationship between Eq. (10) and the experiments looks inconsistent.
  3. [Table 1] Decoder-time reduction is reported as a single 44% without variance or measurement repeats. Since compute reduction is a central contribution, give the same statistical treatment as the quality metrics, e.g., per-sequence or per-run spread and number of repeats.
  4. [Experiments / Native 1080p] State explicitly that the 1080p curves in Fig. 3 use the multi-resolution + predictive-skip variant described in Appendix F, not the pure BPP-to-compute scheduler. Currently a reader may attribute the 1080p gains to the scheduler alone.
  5. [Probe-Free Calibrated Scheduling] The 'strict ultra-low profile' paragraph introduces the symbol eta without defining it; define eta or drop the notation.
  6. [Turbo-DDCM transfer] The abstract and introduction say the same interface covers DDCM-style image compression, but the experiment fixes M=25 and p=7 manually and the paper later notes it tests interface transfer rather than the calibrated schedule. Add a caveat in the main text where the coverage claim is made.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: calibrated constants are labeled as calibration, quality claims are measured, and the schedule mapping is arithmetic.

full rationale

GVCCTurbo's derivation chain is self-contained: the BPP-to-schedule calculator (Eqs. 10-13) is arithmetic once M_perc and tau_gap are fixed, and those two constants are explicitly calibrated, not predicted: 'The cliffsweep that fixes tau_gap for our instantiation is reported in the ablations; the scheduler then uses M_perc and tau_gap to allocate bitrate and compute.' Table 1 reports measured LPIPS/PSNR and decoder wall-clock reductions, so the central quality claim is not the output of a fitted model. The only load-bearing prior result is the GVCC codebook-correction interface, which is cited as prior work (Zeng et al. 2026) and is not constructed from the present paper's outputs. The paper also includes a held-out calibration check (D.2) using clips disjoint from all benchmarks and explicitly disclaims guarantees ('This is local intuition rather than a guarantee'; 'The actual M_perc is therefore calibrated empirically'; 'The current constants are calibrated per protocol and are not shown resolution- or GOP-invariant'). The reviewer's in-sample-calibration concern is an external-validity issue about whether Table 1 constants were selected on the same datasets, not a circular derivation: no equation reduces a claimed result to its own input. Minor self-citation to the authors' own GVCC is present but non-load-bearing for the scheduler contribution.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The scheduler's contribution is an allocation rule plus an empirical cache. It depends on two calibrated constants (M_perc, tau_gap), a deterministic tail length, and several protocol assumptions about path convention, the codebook channel, endpoint staleness, and the dominance of prior-evaluation cost. No new physical or model entities are introduced; the method is zero-training by construction.

free parameters (3)
  • M_perc = 48 (Wan-GVCC 720p protocol)
    Per-step atom count selected from iso-NFE, iso-B0 LPIPS sweeps (Table 2, Appendix D.2). Controls correction strength and per-step bitrate; central to every schedule produced by the calculator.
  • tau_gap = ~0.15
    Skip-gap ratio calibrated from fixed-(T,M) cliff sweeps (Table 3, Appendix D.2). Determines refresh period via p* = floor(tau_gap T)+1; the 'probe-free' mapping depends on this fitted constant.
  • q (deterministic tail length) = 3 for video, 1 for image
    Chosen protocol constant for the number of deterministic no-codebook trailing steps. Enters T* = N_code* + q and N_FE*, affecting the computed schedule and compute count.
assumptions (5)
  • standard math Linear-path rectified flow convention x_t = (1-t)x0 + t*epsilon, with the model predicting epsilon - x0.
    Used in Eq. (2) and Eq. (6) to identify the endpoint residual with the missing noise innovation. If a different path convention is used, a path-specific scalar is needed, as acknowledged for DDPM.
  • domain assumption Top-M signed-atom codebook projection, normalized to sampler scale, is an adequate fixed-rate directional quantization of the innovation.
    The paper argues this is analogous to DiffC's innovation channel, but the sufficiency of a fixed-scale direction-only channel for maintaining reconstruction quality is assumed, not proven.
  • domain assumption Endpoint staleness remains small relative to the residual over the calibrated skip window; discrete Top-M selection can change discontinuously.
    Central approximation. The paper provides local second-order intuition but relies on the empirically calibrated tau_gap for the safety of the skip window (Section 'Endpoint-Cached Prior Refresh').
  • domain assumption Prior evaluation cost dominates per-step codebook correction cost, so N_FE tracks wall-clock time.
    Explicitly stated as a limitation in the Conclusion and Appendix D.4. The image-domain Turbo-DDCM transfer deliberately tests the boundary where this assumption fails.
  • domain assumption Encoder and decoder share a fixed schedule and derive identical refresh steps, so skipped evaluations require no signaling.
    Requires deterministic bit-exact replay. Content-adaptive refresh would require signaling and is explicitly excluded from the method.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GVCCTurbo: Rate-Compute Quality Scheduling for Codebook Driven Generative Compression." pith.science (2026). https://pith.science/paper/ETDXGAQS

@misc{pith2026260803517,
  author       = {Pith},
  title        = {Pith review of: GVCCTurbo: Rate-Compute Quality Scheduling for Codebook Driven Generative Compression},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ETDXGAQS}},
  note         = {Machine review of arXiv:2608.03517}
}
abstract

Codebook-driven generative compression uses a pretrained image or video generator as a zero-shot visual prior and transmits compact codebook indices to guide reconstruction at ultra-low bitrate. Current codecs tie each finite-rate correction to a fresh prior evaluation, so shortening the sampler also removes correction slots that carry target-dependent information. We propose GVCCTurbo, a BPP-driven scheduler that separates expensive prior refreshes from codebook corrections: after calibrating an atom-count operating point and skip-gap ratio once per protocol, it maps a target codebook-payload bitrate to a trajectory length and refresh period, making BPP a schedule input instead of a fixed consequence of sampler length. The same endpoint-prediction and finite-rate steering interface covers GVCC-style rectified-flow video and DDCM-style diffusion image compression, preserving zero-training deployment and compatibility with future distilled priors. Native 1080p curves position the complete zero-shot codec in the ultra-low-bitrate regime. In a controlled 720p Wan-GVCC study, the scheduler cuts prior evaluations from 20 to 9 for a $\sim\!44\%$ measured decoding-time reduction shared across the whole schedule family, at a small shared LPIPS cost on high-motion content; within that family, uniform refresh thinning (pure-skip) is a boundary point, and the BPP-aware interior point trades $2.9\%$ fewer codebook-payload bits for consistently higher PSNR at comparable LPIPS. These results support BPP-to-compute scheduling as a controllable extension of sampler-length tuning, without requiring the allocated point to dominate every boundary point.

Figures

Figures reproduced from arXiv: 2608.03517 by the authors.

Figure 1
Figure 1. Trajectory geometry. (a) Full GVCC spends one [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Rate–compute scheduling. GVCC ties each code [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Native 1080p rate–distortion comparison on HEVC-B, UVG, and MCL-JCV benchmarks under the same metrics and evaluation protocol. BPP is the protocol-total rate. Experimental Protocol For the native 1080p comparison we use UVG (Mercat, Viita￾nen, and Vanne 2020), HEVC Class B, and MCL-JCV (Wang et al. 2016) with 33-frame GOPs; these system-level points use the FLF2V-14B GVCC backbone with predictive skip and multi-reso… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Qualitative 1080p comparison with representative existing methods. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Cross-backbone transfer to FLF2V-14B on UVG. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Probe-free codebook-payload bitrate curve at Mperc=48 on a one-GOP diagnostic subset. Two se￾quences spanning textured-hard (ShakeNDry) to smooth (videoSRC01) share the calculator’s schedules at each bitrate target. This figure supplements the fixed-B0, three-dataset c…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

48 extracted references · 34 canonical work pages

  1. [1]

    Zeng, Ziyue and Su, Xun and Liu, Haoyuan and Lu, Bingyu and Tatsumi, Yui and Watanabe, Hiroshi , journal =

    work page

  2. [2]

    Vaisman, Amit and Ohayon, Guy and Manor, Hila and Elad, Michael and Michaeli, Tomer , booktitle =. Turbo-. 2026 , note =

    work page 2026

  3. [3]

    arXiv preprint arXiv:2503.20314 , year =

    Wan: Open and Advanced Large-Scale Video Generative Models , author =. arXiv preprint arXiv:2503.20314 , year =

    arXiv

  4. [4]

    and Ohm, Jens-Rainer and Han, Woo-Jin and Wiegand, Thomas , journal =

    Sullivan, Gary J. and Ohm, Jens-Rainer and Han, Woo-Jin and Wiegand, Thomas , journal =. Overview of the High Efficiency Video Coding (

    work page

  5. [5]

    and Ohm, Jens-Rainer , journal =

    Bross, Benjamin and Wang, Ye-Kui and Ye, Yan and Liu, Shan and Chen, Jianle and Sullivan, Gary J. and Ohm, Jens-Rainer , journal =. Overview of the Versatile Video Coding (

    work page

  6. [6]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    Neural Video Compression with Feature Modulation , author =. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    work page

  7. [7]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    Towards Practical Real-Time Neural Video Compression , author =. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    work page

  8. [8]

    Generative Latent Coding for Ultra-Low Bitrate Image and Video Compression

    Generative Latent Coding for Ultra-Low Bitrate Image and Video Compression , author =. arXiv preprint arXiv:2505.16177 , year =

    work page Pith review arXiv

Show all 48 references
  1. [9]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    Generative Neural Video Compression via Video Diffusion Prior , author =. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

  2. [10]

    Ling, Xiaoyue and Zhou, Chuqin and Li, Chunyi and Chen, Yunuo and Tian, Yuan and Lu, Guo and Zhang, Wenjun , journal =. Free-

  3. [11]

    arXiv preprint arXiv:2510.09987 , year =

    Generative Latent Video Compression , author =. arXiv preprint arXiv:2510.09987 , year =

  4. [12]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    Generative Video Compression with One-Dimensional Latent Representation , author =. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

  5. [13]

    Li, Daowen and Dong, Ruixiao and Chen, Ying and Li, Kai and Ding, Ding and Li, Li , journal =

  6. [14]

    arXiv preprint arXiv:2512.00408 , year =

    Low-Bitrate Video Compression through Semantic-Conditioned Diffusion , author =. arXiv preprint arXiv:2512.00408 , year =

    arXiv

  7. [15]

    arXiv preprint arXiv:2604.06655 , year =

    Controllable Generative Video Compression , author =. arXiv preprint arXiv:2604.06655 , year =

    arXiv

  8. [16]

    2026 , doi =

    Wang, Zhitao and Man, Hengyu and Li, Wenrui and Wang, Xingtao and Fan, Xiaopeng and Zhao, Debin , booktitle =. 2026 , doi =

    2026

  9. [17]

    2025 , note =

    Wan, Rui and Zheng, Qi and Fan, Yibo , booktitle =. 2025 , note =

    2025

  10. [18]

    Mercat, Alexandre and Viitanen, Marko and Vanne, Jarno , booktitle =

  11. [19]

    Jay , booktitle =

    Wang, Haiqiang and Gan, Weihao and Hu, Sudeng and Lin, Joe Yuchieh and Jin, Lina and Song, Longguang and Wang, Ping and Katsavounidis, Ioannis and Aaron, Anne and Kuo, C.-C. Jay , booktitle =

  12. [20]

    2026 , note =

    Spectral Progressive Diffusion , author =. 2026 , note =

    2026

  13. [21]

    Journal of Machine Learning Research , volume =

    Cascaded Diffusion Models for High Fidelity Image Generation , author =. Journal of Machine Learning Research , volume =

  14. [22]

    International Conference on Learning Representations (ICLR) , year =

    Podell, Dustin and English, Zion and Lacey, Kyle and Blattmann, Andreas and Dockhorn, Tim and M. International Conference on Learning Representations (ICLR) , year =

  15. [23]

    International Conference on Machine Learning (ICML) , year =

    Scaling Rectified Flow Transformers for High-Resolution Image Synthesis , author =. International Conference on Machine Learning (ICML) , year =

  16. [24]

    International Conference on Learning Representations (ICLR) , year =

    Progressive Distillation for Fast Sampling of Diffusion Models , author =. International Conference on Learning Representations (ICLR) , year =

  17. [25]

    International Conference on Machine Learning (ICML) , year =

    Consistency Models , author =. International Conference on Machine Learning (ICML) , year =

  18. [26]

    2023 , note =

    Latent Consistency Models: Synthesizing High-Resolution Images with Few-Step Inference , author =. 2023 , note =

    2023

  19. [27]

    Xie, Enze and Chen, Junsong and Chen, Junyu and Cai, Han and Tang, Haotian and Lin, Yujun and Zhang, Zhekai and Li, Muyang and Zhu, Ligeng and Lu, Yao and Han, Song , year =

  20. [28]

    Ma, Xinyin and Fang, Gongfan and Wang, Xinchao , booktitle =

  21. [29]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    Timestep Embedding Tells: It's Time to Cache for Video Diffusion Model , author =. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

  22. [30]

    2025 , note =

    Ma, Zehong and Wei, Longhui and Wang, Feng and Zhang, Shiliang and Tian, Qi , booktitle =. 2025 , note =

    2025

  23. [31]

    2026 , note =

    Haghighi, Yasaman and Alahi, Alexandre , booktitle =. 2026 , note =

    2026

  24. [32]

    Fan, Zhentao and Wang, Zongzuo and Zhang, Weiwei , journal =

  25. [33]

    2026 , note =

    Bu, Jiazi and Ling, Pengyang and Zhou, Yujie and Wang, Yibin and Zang, Yuhang and Wu, Tong and Lin, Dahua and Wang, Jiaqi , booktitle =. 2026 , note =

    2026

  26. [34]

    Advances in Neural Information Processing Systems , year =

    High-Fidelity Generative Image Compression , author =. Advances in Neural Information Processing Systems , year =

  27. [35]

    arXiv preprint arXiv:2206.08889 , year =

    Lossy Compression with Gaussian Diffusion , author =. arXiv preprint arXiv:2206.08889 , year =

    arXiv

  28. [36]

    Advances in Neural Information Processing Systems , year =

    Lossy Image Compression with Conditional Diffusion Models , author =. Advances in Neural Information Processing Systems , year =

  29. [37]

    International Conference on Learning Representations , year =

    Lossy Compression with Pretrained Diffusion Models , author =. International Conference on Learning Representations , year =

  30. [38]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    The Unreasonable Effectiveness of Deep Features as a Perceptual Metric , author =. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

  31. [39]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , volume =

    Image Quality Assessment: Unifying Structure and Texture Similarity , author =. IEEE Transactions on Pattern Analysis and Machine Intelligence , volume =

  32. [40]

    International Conference on Machine Learning (ICML) , year =

    Compressed Image Generation with Denoising Diffusion Codebook Models , author =. International Conference on Machine Learning (ICML) , year =

  33. [41]

    Advances in Neural Information Processing Systems , year=

    Denoising Diffusion Probabilistic Models , author=. Advances in Neural Information Processing Systems , year=

  34. [42]

    International Conference on Learning Representations (ICLR) , year=

    Score-Based Generative Modeling through Stochastic Differential Equations , author=. International Conference on Learning Representations (ICLR) , year=

  35. [43]

    International Conference on Learning Representations (ICLR) , year=

    Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow , author=. International Conference on Learning Representations (ICLR) , year=

  36. [44]

    arXiv preprint arXiv:2407.09896 , year=

    Zero-Shot Image Compression with Diffusion-Based Posterior Sampling , author=. arXiv preprint arXiv:2407.09896 , year=

    arXiv

  37. [45]

    arXiv preprint arXiv:2510.23633 , year=

    Noise is All You Need: Solving Linear Inverse Problems by Noise Combination Sampling with Diffusion Models , author=. arXiv preprint arXiv:2510.23633 , year=

    arXiv

  38. [46]

    arXiv preprint arXiv:2601.18932 , year=

    Advances in Diffusion-Based Generative Compression , author=. arXiv preprint arXiv:2601.18932 , year=

  39. [47]

    arXiv preprint arXiv:2602.22654 , year=

    Denoising as Path Planning: Training-Free Acceleration of Diffusion Models with DPCache , author=. arXiv preprint arXiv:2602.22654 , year=

  40. [48]

    arXiv preprint arXiv:2507.02860 , year=

    Less is Enough: Training-Free Video Diffusion Acceleration via Runtime-Adaptive Caching , author=. arXiv preprint arXiv:2507.02860 , year=

    arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.