Pith. sign in

REVIEW 4 major objections 5 minor 52 references

GenVC claims the first video diffusion codec trained from scratch, beating billion-scale baselines at ultra-low bitrates.

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

2026-08-01 06:07 UTC pith:XRZVZZ7V

load-bearing objection A genuinely new compression-oriented video diffusion codec with a plausible gating fix for one-step distillation — but the failure-mode story is inferred, not shown, and the baselines are mostly its own. the 4 major comments →

arxiv 2607.22772 v1 pith:XRZVZZ7V submitted 2026-07-24 eess.IV cs.CV

Generative Video Compression with Adaptive Score Distillation

Naifu Xue , Zhaoyang Jia , Haosen Li , Zihan Zheng , Jiahao Li , Bin Li , Xiaoyi Zhang , Qi Meng
show 2 more authors
Yuan Zhang Yan Lu
This is my paper
classification eess.IV cs.CV
keywords generative video compressionvideo diffusion modelone-step distillationdistribution matching distillationadaptive score distillationmotion stallingperceptual qualitypixel-space diffusion
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 tries to establish that a video diffusion model designed and trained specifically for compression can outperform codecs built by adapting billion-scale pretrained generation backbones, at a fraction of the parameters and with one-step decoding. It also claims to diagnose why direct distribution-matching distillation fails for video—producing motion-stalled reconstructions—and to fix this with an alignment gate that weighs each distillation update by its cosine similarity to the ground-truth direction. If true, generative video compression does not need huge pretrained backbones; it can be learned end-to-end for the compression task itself. The reported results include average bitrate savings of 62.5% at matched LPIPS and 71.3% at matched FID over a strong generative baseline, with a 478M-parameter diffusion model decoding 1080p at 15.1 fps.

Core claim

GenVC is a pixel-space video diffusion model with a global-to-local hierarchy (a coarse spatio-temporal DiT followed by patch-level and frame-level refiners), jointly trained with a contextual video codec. When distilling the multi-step model into one step, ordinary DMD drives the student toward reconstructions with stalled or frozen motion. The paper traces this to a teacher-side guidance failure: once student-induced perturbations leave the frozen teacher's training region, the teacher's score predictions become misleading, and the DMD update reinforces rather than corrects the drift. The proposed fix, Adaptive Score Distillation, assigns each training sample a weight from a sigmoid over t

What carries the argument

The central object is the per-sample alignment gate in Adaptive Score Distillation: w_b = sigmoid((cos(g_DMD^(b), X^(b) − X_stu^(b)) − τ)/κ), applied as a stop-gradient weight to the DMD gradient. The gate is what breaks the feedback loop that would otherwise drive the student into motion stalling, by suppressing updates that point away from the ground truth while keeping distribution-matching updates that pull toward realistic detail. The architecture that carries the compression side is the global-to-local pixel-space hierarchy, which recovers fine spatial texture and temporal variation without the memory blow-up of per-pixel modulation at 1080p.

Load-bearing premise

The method assumes that cosine alignment between the DMD update and the ground-truth direction reliably indicates whether the frozen teacher's guidance is corrective; if that correlation is weak, the gate could suppress legitimate stochastic diversity or fail to catch genuinely misleading updates.

What would settle it

During one-step distillation, compute for a large batch the gate weight w against the actual effect of applying the update: if many samples with w > 0.5 yield updates that increase distance from the ground truth (or samples with w < 0.5 would have decreased it), the gate is not tracking correctness. Alternatively, train the same one-step model without the gate but with a teacher whose training-time noise distribution includes student-like perturbations; if motion stalling no longer occurs, the teacher-side failure story would be undermined.

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

If this is right

  • If the claims hold, generative video codecs can be built from scratch with sub-billion-parameter diffusion models, rather than inheriting billion-scale text-to-video backbones.
  • One-step diffusion distillation can be made stable for video reconstruction by using a ground-truth-anchored gate, suggesting a general recipe for compression-oriented distillation.
  • The bitrate savings at matched perceptual quality imply that ultra-low-bitrate streaming could deliver visually realistic detail while transmitting only a few hundred bytes per frame.
  • The joint training of codec and diffusion model means the transmitted representation and the generative decoder co-adapt, potentially reshaping how rate-distortion trade-offs are struck in practice.
  • If the student improves upon its multi-step teacher, it suggests that distillation with source-anchored supervision can surpass the teacher's own sampled reconstructions, not merely approach them.

Where Pith is reading between the lines

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

  • The gate's reliance on ground-truth direction means it is only applicable where the source video is available during training, i.e., compression; it would not transfer directly to unconditional generation, but a similar principle might be adapted to self-supervised settings with pseudo-ground-truth.
  • The teacher-side guidance failure may be a more general phenomenon in score distillation: whenever a student's inputs drift from the teacher's training manifold, the teacher's gradients can become adversarial. This suggests that out-of-distribution detection or input-perturbation checks could be a broader safeguard for distillation.
  • One testable extension is to replace the hand-chosen sigmoid threshold τ and temperature κ with a learned or calibrated gate, or to use a per-pixel or per-timestep alignment measure instead of a full-sample cosine, which might further improve the fidelity-diversity trade-off.
  • The reported gains could be partly due to the underlying DCVC-style codec backbone rather than the diffusion-specific components; an ablation that replaces GenVC's diffusion decoder with a non-diffusion generative head while keeping the same codec would isolate the diffusion contribution.

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 presents GenVC, a generative video codec built on a pixel-space video diffusion model trained from scratch jointly with a contextual codec, rather than adapting a large pretrained text-to-video backbone. A global-to-local architecture (DiT plus patch and frame refiners) is used to keep pixel-space modeling tractable at 1080p. The multi-step model is then distilled into a one-step student with distribution matching distillation (DMD). The authors observe that direct DMD causes motion stalling, which they attribute to a "teacher-side guidance failure": student-induced perturbations leave the frozen teacher's training region, so the teacher's score estimates become misleading. To fix this, they propose Adaptive Score Distillation, a per-sample gate that downweights DMD updates whose cosine alignment with the ground-truth direction is below a threshold. Experiments report large perceptual gains over prior codecs, including average bitrate savings of 62.5% at matched LPIPS and 71.3% at matched FID relative to GLVC, with one-step 1080p decoding at 15.1 fps on an A100 using a 478.0M-parameter diffusion model.

Significance. If the reported results are robust, GenVC is a significant systems contribution: it demonstrates that a compression-specialized, sub-billion-parameter video diffusion model can outperform codecs built on billion-scale pretrained backbones at ultra-low bitrates, and that one-step distillation can be made stable for video reconstruction. The paper's internal evidence is unusually complete for a systems paper: rate-distortion curves on three datasets, a progressive architecture ablation (Fig. 11), a clean distillation ablation (Table II), and a training-dynamics diagnostic (Fig. 9). The proposed adaptive gate is simple and clearly described. The main weaknesses are that the central mechanistic claim (teacher-side OOD failure) is not directly tested, the gate hyperparameters are not sensitivity-analyzed, and the headline comparison is anchored to the authors' own prior work without code or model release for independent verification.

major comments (4)
  1. [Sec. IV-B, Fig. 8-9]
  2. [Sec. IV-C, Eq. (13), Sec. V-A]
  3. [Table II]
  4. [Sec. V-B, Table I]
minor comments (5)
  1. [Abstract, Sec. I]
  2. [Sec. IV-A, Eq. (9)-(10)]
  3. [Sec. III-A, Fig. 4]
  4. [Sec. V-A]
  5. [Sec. V-A, Fig. 10]

Circularity Check

0 steps flagged

No significant circularity: the adaptive gate is a defined training-time supervision signal, and the paper's self-citations are not load-bearing to the central derivation.

full rationale

The central derivation chain is self-contained. The multi-step teacher is trained from scratch via flow matching (Eq. 3) with codec rate/distortion terms; the one-step student is distilled with DMD (Eq. 9), an adaptive alignment gate (Eqs. 12-14), and ground-truth-anchored losses (Eq. 17). The gate uses the ground-truth direction X - Xhat_stu as a training-time supervision signal; it is a defined component of the objective, not a parameter fitted to a metric that is later reported as an independent prediction. The teacher-side OOD diagnosis in Sec. IV-B is a mechanistic hypothesis supported by Figs. 8-9 but not directly measured (no OOD-coverage test); this is an evidentiary gap, not circularity. The paper cites several prior works by the same authors (CoD, S2VC, GLVC, GLC-video, DCVC) and uses GLVC as the BD-Rate anchor, but the comparison is an empirical benchmark against a published system, and the distillation mechanism does not reduce to any of those citations. No equation in the paper is equivalent by construction to its inputs; no fitted parameter is renamed as a prediction. Hence no significant circularity.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 0 invented entities

The central claim rests on standard flow-matching and DMD machinery (axioms 1-2), plus domain assumptions about paired-data availability, metric validity, and resolution generalization. The hand-set hyperparameters (λ_D, τ, κ, CFG scale) are not derived from theory; the headline bitrate-savings numbers are conditional on them. No invented entities.

free parameters (5)
  • Rate-distortion weight λ_R = multi-step: {0.56,0.80,1.12,1.76,2.40}; student: {10.24,5.12,2.56,1.28,0.64,0.32}
    Swept to produce rate points; standard codec practice, but the bitrate operating range of the final comparisons depends on these choices.
  • Loss weights λ_aux, λ_L, λ_F, λ_D = 0.01, 1.0, 2.0, 8.0
    Hand-chosen with no sensitivity analysis. λ_F weights the RAFT-flow distortion that directly suppresses motion error, partially overlapping with the gate's effect.
  • Gate threshold τ and temperature κ = τ=0.015, κ=0.02
    Hand-set (Sec. V-A); the central gating mechanism's behavior and the reported BD-Rate/BD-ME numbers depend on these values, with no ablation across them.
  • Teacher CFG scale = 1.5
    Used for teacher real-score estimates and 20-step sampling; DMD's quality and the 'real score' interpretation depend on this.
  • Diffusion-time sampling for student DMD = sigmoid(Gaussian) rescaled to [0.05, 0.95]
    A design choice for the DMD noise levels; affects which perturbations the gate sees.
axioms (6)
  • standard math Flow-matching/rectified-flow path X_t = (1−t)X_0 + tX_1 with conditional velocity X_1 − X_0 (Eq. 1) and the clean-prediction-to-velocity conversion (Eq. 2).
    Invoked without proof in Sec. III-A; the entire diffusion training and DMD real/fake score interpretation rest on it.
  • standard math DMD gradient direction g_DMD = X̂_real − X̂_fake (Eq. 10) approximates ∇_θ KL between student and real distributions.
    Taken from Yin et al. [16]; the paper's gating modifies exactly this quantity, so the derivation's validity is inherited.
  • domain assumption Paired ground-truth videos X are available during codec training and can be used per-sample in the gate (Eq. 12).
    True in compression training, but it means the gate cannot be applied at inference; the student must learn from gated gradients alone. Explicit in Sec. IV-C.
  • domain assumption The frozen teacher's CFG-scale-1.5 prediction on off-distribution student inputs behaves as a 'real' score that may become misleading; training-region coverage is not measured.
    Sec. IV-B states the failure mechanism but does not directly quantify out-of-training-distribution inputs; the whole Adaptive Score Distillation story rests on this.
  • domain assumption Perceptual metrics (LPIPS, DISTS, FID, FloLPIPS) are valid proxies for quality at ultra-low bitrates, and BD-Rate at 'matched LPIPS/FID' is a fair comparison across codecs with different rate ranges.
    Sec. V-A/V-B define the headline savings through these metrics; if the metrics reward generative smoothing, the 'SOTA perceptual quality' claim overstates user-perceived quality.
  • domain assumption Training at 512×512 crops from OpenVid generalizes to 1080p test sequences.
    Sec. V-A: models train on 512×512 clips, evaluate at 1920×1080; standard in neural codecs but a resolution generalization assumption.

pith-pipeline@v1.3.0-alltime-deepseek · 18079 in / 17163 out tokens · 158214 ms · 2026-08-01T06:07:49.578524+00:00 · methodology

0 comments
read the original abstract

Diffusion models provide strong generative capabilities for video compression at ultra-low bitrates. Existing diffusion-based video codecs adapt base models originally developed for text-conditioned generation, whereas diffusion models designed and trained specifically for compression remain unexplored. To fill this gap, we introduce our Generative Video Codec (GenVC), built on a video diffusion model trained from scratch for compression. To our knowledge, this is the first compression-oriented video diffusion model. We realize this model directly in pixel space with a global-to-local hierarchy that recovers fine spatio-temporal details, enabling high-quality generative reconstruction from compressed representations. To accelerate inference, we distill the multi-step model into one step using distribution matching distillation (DMD). Applying DMD directly, however, drives the student toward motion-stalled reconstructions. We trace this to a teacher-side guidance failure: once student-induced perturbations leave the frozen teacher's training region, its guidance can become misleading, causing DMD updates to reinforce rather than correct the student drift. To break the resulting feedback loop, we propose Adaptive Score Distillation, which gates DMD updates according to their alignment with the ground-truth direction, enabling high-quality reconstruction with coherent motion. Experimental results show that GenVC achieves state-of-the-art perceptual quality at ultra-low bitrates, with average bitrate savings of 62.5% at matched LPIPS and 71.3% at matched FID over GLVC. Unlike prior codecs that inherit billion-scale pretrained backbones, our diffusion model has only 478.0M parameters and decodes 1080p video in a single step at 15.1 fps on an A100 GPU.

Figures

Figures reproduced from arXiv: 2607.22772 by Bin Li, Haosen Li, Jiahao Li, Naifu Xue, Qi Meng, Xiaoyi Zhang, Yan Lu, Yuan Zhang, Zhaoyang Jia, Zihan Zheng.

Figure 1
Figure 1. Figure 1: Overview of GenVC. The multi-step model (top) serves as the teacher [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Frame examples. GenVC delivers the best perceptual quality at the lowest bitrate among the compared methods. Zoom in for better visualization. [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Adaptive gating prevents motion stalling. Horizontal strips extracted from consecutive frames are stacked from top to bottom, illustrating the motion trajectory over time. Close agreement with the GT strip indicates faithful motion reconstruction, whereas pronounced staircase patterns reveal motion stalling and abrupt jumps between adjacent frames. been demonstrated in image generation and compression [12]… view at source ↗
Figure 4
Figure 4. Figure 4: Overview of the GenVC pipeline. The clean source GOP and its decoded condition sequence are denoted by X0 ≡ X = (x (1), . . . , x(K) ) and c = (c (1), . . . , c(K) ), respectively. Q denotes quantization; AE and AD denote the arithmetic encoder and decoder. on compression-specific detail synthesis. This specialization allows GenVC to achieve SOTA perceptual quality at ultra￾low bitrates with a sub-billion-… view at source ↗
Figure 5
Figure 5. Figure 5: Global-to-local pixel-space architecture. A coarse spatio-temporal DiT [PITH_FULL_IMAGE:figures/full_fig_p003_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Progressive architecture ablation on HEVC-B, reported as BD-LPIPS [PITH_FULL_IMAGE:figures/full_fig_p004_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Teacher-side guidance failure and Adaptive Score Distillation. Left: vanilla DMD produces the update gDMD. Middle: the green area denotes the region covered during teacher training. When student reconstructions remain close to this region, teacher guidance yields stable, corrective DMD updates; substantial deviations may instead yield misleading updates. Right: our alignment-based gate attenuates such misl… view at source ↗
Figure 8
Figure 8. Figure 8: Visual evidence of the teacher-side guidance failure. Optical flow of the ground truth and the stable/stalled student reconstructions. Perturbed versions of both student outputs are passed through the same frozen teacher: motion from the stable student is largely preserved, whereas the already degraded motion of the stalled student is suppressed further. student distribution ( [PITH_FULL_IMAGE:figures/ful… view at source ↗
Figure 9
Figure 9. Figure 9: Training dynamics of the teacher-side guidance failure. Training￾time cosine (solid, left axis) and evaluation-time Motion Error (dashed, right axis) for the stalled student (red) and the stable student (green). ground truth. To break the feedback loop, we turn this signal into a per-sample gate: each sample b in a batch has its update weighted by a smooth function of its alignment, wb = σ [PITH_FULL_IMAG… view at source ↗
Figure 10
Figure 10. Figure 10: Rate–quality performance on HEVC Class B, UVG, and MCL-JCV. The multi-step model is evaluated as a standalone codec, while GenVC denotes [PITH_FULL_IMAGE:figures/full_fig_p008_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Architecture ablation of the multi-step teacher, with the parameter [PITH_FULL_IMAGE:figures/full_fig_p009_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

52 extracted references · 9 linked inside Pith

  1. [1]

    Perceptual learned video compression with recurrent conditional gan

    R. Yang, R. Timofte, and L. Van Gool, “Perceptual learned video compression with recurrent conditional gan.” inIJCAI, 2022, pp. 1537– 1544

    2022

  2. [2]

    Diffusion-based perceptual neural video com- pression with temporal diffusion information reuse,

    W. Ma and Z. Chen, “Diffusion-based perceptual neural video com- pression with temporal diffusion information reuse,”arXiv preprint arXiv:2501.13528, 2025

    Pith/arXiv arXiv 2025

  3. [3]

    Generative latent coding for ultra-low bitrate image and video compression,

    L. Qi, Z. Jia, J. Li, B. Li, H. Li, and Y . Lu, “Generative latent coding for ultra-low bitrate image and video compression,”IEEE Transactions on Circuits and Systems for Video Technology, 2025

    2025

  4. [4]

    Open-sora: Democratizing efficient video production for all,

    Z. Zheng, X. Peng, T. Yang, C. Shen, S. Li, H. Liu, Y . Zhou, T. Li, and Y . You, “Open-sora: Democratizing efficient video production for all,” arXiv preprint arXiv:2412.20404, 2024

    Pith/arXiv arXiv 2024

  5. [5]

    Cogvideox: Text-to-video diffusion models with an expert transformer,

    Z. Yang, J. Teng, W. Zheng, M. Ding, S. Huang, J. Xu, Y . Yang, W. Hong, X. Zhang, G. Feng, D. Yin, Y . Zhang, W. Wang, Y . Cheng, B. Xu, X. Gu, Y . Dong, and J. Tang, “Cogvideox: Text-to-video diffusion models with an expert transformer,” inInternational Conference on Learning Representations, 2025

    2025

  6. [6]

    Hunyuanvideo: A systematic framework for large video generative models,

    W. Kong, Q. Tian, Z. Zhang, R. Min, Z. Dai, J. Zhou, J. Xiong, X. Li, B. Wu, J. Zhanget al., “Hunyuanvideo: A systematic framework for large video generative models,”arXiv preprint arXiv:2412.03603, 2024

    Pith/arXiv arXiv 2024

  7. [7]

    Wan: Open and advanced large-scale video generative models,

    T. Wan, A. Wang, B. Ai, B. Wen, C. Mao, C.-W. Xie, D. Chen, F. Yu, H. Zhao, J. Yanget al., “Wan: Open and advanced large-scale video generative models,”arXiv preprint arXiv:2503.20314, 2025

    Pith/arXiv arXiv 2025

  8. [8]

    Generative neural video compression via video diffusion prior,

    Q. Mao, H. Cheng, T. Yang, L. Jin, and S. Ma, “Generative neural video compression via video diffusion prior,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2026, pp. 43 239–43 248

    2026

  9. [9]

    Gvcc: Zero-shot video compression via codebook-driven stochastic rectified flow,

    Z. Zeng, X. Su, H. Liu, B. Lu, Y . Tatsumi, and H. Watanabe, “Gvcc: Zero-shot video compression via codebook-driven stochastic rectified flow,”arXiv preprint arXiv:2603.26571, 2026. 10

    Pith/arXiv arXiv 2026

  10. [10]

    Efficient video diffusion with sparse information transmission for video compression,

    M. Zhou, Z. Chen, and Y . Zhang, “Efficient video diffusion with sparse information transmission for video compression,”arXiv preprint arXiv:2603.18501, 2026

    arXiv 2026

  11. [11]

    Cod: A diffusion foundation model for image compression,

    Z. Jia, Z. Zheng, N. Xue, J. Li, B. Li, Z. Guo, X. Zhang, H. Li, and Y . Lu, “Cod: A diffusion foundation model for image compression,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2026, pp. 38 420–38 429

    2026

  12. [12]

    Cod-lite: Real-time diffusion- based generative image compression,

    Z. Jia, N. Xue, Z. Zheng, J. Li, B. Li, X. Zhang, Z. Guo, Y . Zhang, H. Li, and Y . Lu, “Cod-lite: Real-time diffusion- based generative image compression,” 2026. [Online]. Available: https://arxiv.org/abs/2604.12525

    Pith/arXiv arXiv 2026

  13. [13]

    Diffvc-osd: One-step diffusion-based perceptual neural video compression framework,

    W. Ma and Z. Chen, “Diffvc-osd: One-step diffusion-based perceptual neural video compression framework,”arXiv preprint arXiv:2508.07682, 2025

    Pith/arXiv arXiv 2025

  14. [14]

    Diffvc-rt: Towards practical real-time diffusion-based perceptual neural video compression,

    ——, “Diffvc-rt: Towards practical real-time diffusion-based perceptual neural video compression,”arXiv preprint arXiv:2601.20564, 2026

    arXiv 2026

  15. [15]

    Single- step diffusion-based video coding with semantic-temporal guidance,

    N. Xue, Z. Jia, J. Li, B. Li, Z. Zheng, Y . Zhang, and Y . Lu, “Single- step diffusion-based video coding with semantic-temporal guidance,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2026, pp. 9752–9761

    2026

  16. [16]

    One-step diffusion with distribution matching distillation,

    T. Yin, M. Gharbi, R. Zhang, E. Shechtman, F. Durand, W. T. Freeman, and T. Park, “One-step diffusion with distribution matching distillation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2024, pp. 6613–6623

    2024

  17. [17]

    Improved distribution matching distillation for fast image synthesis,

    T. Yin, M. Gharbi, T. Park, R. Zhang, E. Shechtman, F. Durand, and W. T. Freeman, “Improved distribution matching distillation for fast image synthesis,” inNeurIPS, 2024

    2024

  18. [18]

    One-step diffusion-based image compression with semantic distillation,

    N. Xue, Z. Jia, J. Li, B. Li, Y . Zhang, and Y . Lu, “One-step diffusion-based image compression with semantic distillation,” in Advances in Neural Information Processing Systems, D. Belgrave, C. Zhang, H. Lin, R. Pascanu, P. Koniusz, M. Ghassemi, and N. Chen, Eds., vol. 38. Curran Associates, Inc., 2025, pp. 37 108–37 144. [Online]. Available: https://pr...

    2025

  19. [19]

    The unreasonable effectiveness of deep features as a perceptual metric,

    R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang, “The unreasonable effectiveness of deep features as a perceptual metric,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018, pp. 586–595

    2018

  20. [20]

    Generative latent video compression,

    Z. Guo, Z. Jia, J. Li, X. Zhang, B. Li, and Y . Lu, “Generative latent video compression,”arXiv preprint arXiv:2510.09987, 2025

    arXiv 2025

  21. [21]

    Deep contextual video compression,

    J. Li, B. Li, and Y . Lu, “Deep contextual video compression,” in Advances in Neural Information Processing Systems, vol. 34, 2021, pp. 18 114–18 125

    2021

  22. [22]

    Neural video compression with diverse contexts,

    ——, “Neural video compression with diverse contexts,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion (CVPR), June 2023, pp. 22 616–22 626

    2023

  23. [23]

    Neural video compression with feature modulation,

    ——, “Neural video compression with feature modulation,” inProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2024, pp. 26 099–26 108

    2024

  24. [24]

    Towards practical real-time neural video compression,

    Z. Jia, B. Li, J. Li, W. Xie, L. Qi, H. Li, and Y . Lu, “Towards practical real-time neural video compression,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2025, pp. 12 543–12 552

    2025

  25. [25]

    Ultra-fast neural video compression,

    J. Li, W. Xie, Z. Jia, B. Li, Z. Guo, X. Zhang, and Y . Lu, “Ultra-fast neural video compression,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2026, pp. 41 311–41 321

    2026

  26. [26]

    Overview of the versatile video coding (vvc) standard and its applications,

    B. Bross, Y .-K. Wang, Y . Ye, S. Liu, J. Chen, G. J. Sullivan, and J.- R. Ohm, “Overview of the versatile video coding (vvc) standard and its applications,”IEEE Transactions on Circuits and Systems for Video Technology, vol. 31, no. 10, pp. 3736–3764, 2021

    2021

  27. [27]

    Neural video compression using gans for detail synthe- sis and propagation,

    F. Mentzer, E. Agustsson, J. Ball ´e, D. Minnen, N. Johnston, and G. Toderici, “Neural video compression using gans for detail synthe- sis and propagation,” inEuropean Conference on Computer Vision. Springer, 2022, pp. 562–578

    2022

  28. [28]

    High- resolution image synthesis with latent diffusion models,

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer, “High- resolution image synthesis with latent diffusion models,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion (CVPR), June 2022, pp. 10 684–10 695

    2022

  29. [29]

    Yoda: Yet another one-step diffusion-based video compressor,

    X. Li, J. Zhang, J. Shi, M. Lu, and Z. Ma, “Yoda: Yet another one-step diffusion-based video compressor,”arXiv preprint arXiv:2601.01141, 2026

    arXiv 2026

  30. [30]

    Back to basics: Let denoising generative models denoise,

    T. Li and K. He, “Back to basics: Let denoising generative models denoise,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2026, pp. 36 115–36 125

    2026

  31. [31]

    Pixnerd: Pixel neural field diffusion,

    S. Wang, Z. Gao, C. Zhu, W. Huang, and L. Wang, “Pixnerd: Pixel neural field diffusion,” inThe Fourteenth International Conference on Learning Representations, 2026. [Online]. Available: https: //openreview.net/forum?id=BDnOrExHmt

    2026

  32. [32]

    Dip: Taming diffusion models in pixel space,

    Z. Chen, J. Zhu, X. Chen, J. Zhang, X. Hu, H. Zhao, C. Wang, J. Yang, and Y . Tai, “Dip: Taming diffusion models in pixel space,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2026, pp. 36 136–36 146

    2026

  33. [33]

    Deco: Frequency- decoupled pixel diffusion for end-to-end image generation,

    Z. Ma, L. Wei, S. Wang, S. Zhang, and Q. Tian, “Deco: Frequency- decoupled pixel diffusion for end-to-end image generation,” inPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2026, pp. 43 600–43 610

    2026

  34. [34]

    Pixeldit: Pixel diffusion transformers for image generation,

    Y . Yu, W. Xiong, W. Nie, Y . Sheng, S. Liu, and J. Luo, “Pixeldit: Pixel diffusion transformers for image generation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2026, pp. 14 273–14 282

    2026

  35. [35]

    Adaptive 1d video diffusion autoencoder,

    Y . Teng, M. Lin, X. Liu, S. Wang, X. Yang, and X. Liu, “Adaptive 1d video diffusion autoencoder,” 2026. [Online]. Available: https://arxiv.org/abs/2602.04220

    arXiv 2026

  36. [36]

    Score distillation sampling with learned manifold corrective,

    T. Alldieck, N. Kolotouros, and C. Sminchisescu, “Score distillation sampling with learned manifold corrective,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 1–18

    2024

  37. [37]

    Prolific- dreamer: High-fidelity and diverse text-to-3d generation with variational score distillation,

    Z. Wang, C. Lu, Y . Wang, F. Bao, C. Li, H. Su, and J. Zhu, “Prolific- dreamer: High-fidelity and diverse text-to-3d generation with variational score distillation,”Advances in neural information processing systems, vol. 36, pp. 8406–8441, 2023

    2023

  38. [38]

    From slow bidirectional to fast autoregressive video diffusion models,

    T. Yin, Q. Zhang, R. Zhang, W. T. Freeman, F. Durand, E. Shechtman, and X. Huang, “From slow bidirectional to fast autoregressive video diffusion models,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025, pp. 22 963–22 974

    2025

  39. [39]

    Flow matching for generative modeling,

    Y . Lipman, R. T. Q. Chen, H. Ben-Hamu, M. Nickel, and M. Le, “Flow matching for generative modeling,” inThe Eleventh International Conference on Learning Representations, 2023. [Online]. Available: https://openreview.net/forum?id=PqvMRDCJT9t

    2023

  40. [40]

    Flow straight and fast: Learning to generate and transfer data with rectified flow,

    X. Liu, C. Gong, and Q. Liu, “Flow straight and fast: Learning to generate and transfer data with rectified flow,” inThe Eleventh International Conference on Learning Representations, 2023. [Online]. Available: https://openreview.net/forum?id=XVjTT1nw5z

    2023

  41. [41]

    Raft: Recurrent all-pairs field transforms for optical flow,

    Z. Teed and J. Deng, “Raft: Recurrent all-pairs field transforms for optical flow,” inECCV, 2020, pp. 402–419

    2020

  42. [42]

    Decoupled weight decay regularization,

    I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,” arXiv preprint arXiv:1711.05101, 2017

    Pith/arXiv arXiv 2017

  43. [43]

    Openvid-1m: A large-scale high-quality dataset for text-to-video generation,

    K. Nan, R. Xie, P. Zhou, T. Fan, Z. Yang, Z. Chen, X. Li, J. Yang, and Y . Tai, “Openvid-1m: A large-scale high-quality dataset for text-to-video generation,”arXiv preprint arXiv:2407.02371, 2024

    Pith/arXiv arXiv 2024

  44. [44]

    Uvg dataset: 50/120fps 4k sequences for video codec analysis and development,

    A. Mercat, M. Viitanen, and J. Vanne, “Uvg dataset: 50/120fps 4k sequences for video codec analysis and development,” inProceedings of the 11th ACM multimedia systems conference, 2020, pp. 297–302

    2020

  45. [45]

    Mcl-jcv: a jnd-based h. 264/avc video quality assessment dataset,

    H. Wang, W. Gan, S. Hu, J. Y . Lin, L. Jin, L. Song, P. Wang, I. Katsavounidis, A. Aaron, and C.-C. J. Kuo, “Mcl-jcv: a jnd-based h. 264/avc video quality assessment dataset,” in2016 IEEE international conference on image processing (ICIP). IEEE, 2016, pp. 1509–1513

    2016

  46. [46]

    Overview of the high efficiency video coding (hevc) standard,

    G. J. Sullivan, J.-R. Ohm, W.-J. Han, and T. Wiegand, “Overview of the high efficiency video coding (hevc) standard,”IEEE Transactions on circuits and systems for video technology, vol. 22, no. 12, pp. 1649– 1668, 2012

    2012

  47. [47]

    Free- gvc: Towards training-free extreme generative video compression with temporal coherence,

    X. Ling, C. Zhou, C. Li, Y . Chen, Y . Tian, G. Lu, and W. Zhang, “Free- gvc: Towards training-free extreme generative video compression with temporal coherence,”arXiv preprint arXiv:2602.09868, 2026

    arXiv 2026

  48. [48]

    Image quality assessment: Unifying structure and texture similarity,

    K. Ding, K. Ma, S. Wang, and E. P. Simoncelli, “Image quality assessment: Unifying structure and texture similarity,”IEEE transactions on pattern analysis and machine intelligence, vol. 44, no. 5, pp. 2567– 2581, 2020

    2020

  49. [49]

    Gans trained by a two time-scale update rule converge to a local nash equilibrium,

    M. Heusel, H. Ramsauer, T. Unterthiner, B. Nessler, and S. Hochreiter, “Gans trained by a two time-scale update rule converge to a local nash equilibrium,”Advances in neural information processing systems, vol. 30, 2017

    2017

  50. [50]

    Flolpips: A bespoke video quality metric for frame interpolation,

    D. Danier, F. Zhang, and D. Bull, “Flolpips: A bespoke video quality metric for frame interpolation,” in2022 Picture Coding Symposium (PCS). IEEE, 2022, pp. 283–287

    2022

  51. [51]

    Multiscale structural similarity for image quality assessment,

    Z. Wang, E. Simoncelli, and A. Bovik, “Multiscale structural similarity for image quality assessment,” inThe Thirty-Seventh Asilomar Confer- ence on Signals, Systems & Computers, 2003, vol. 2, 2003, pp. 1398– 1402

    2003

  52. [52]

    Calculation of average psnr differences between rd- curves,

    G. Bjontegaard, “Calculation of average psnr differences between rd- curves,”ITU SG16 Doc. VCEG-M33, 2001

    2001