pith. sign in

arxiv: 2605.30351 · v1 · pith:ZCMI53QInew · submitted 2026-05-28 · 💻 cs.CV · cs.AI

VideoMLA: Low-Rank Latent KV Cache for Minute-Scale Autoregressive Video Diffusion

Hidir Yesiltepe , Jiazhen Hu , Tuna Han Salih Meral , Adil Kaan Akan , Kaan Oktay , Hoda Eldardiry , Pinar Yanardag This is my paper

Pith reviewed 2026-06-29 08:15 UTC · model grok-4.3

classification 💻 cs.CV cs.AI
keywords video diffusionKV cache compressionlow-rank attentionautoregressive video generationMulti-Head Latent Attention3D-RoPEmemory efficiencylong-horizon generation
0
0 comments X

The pith

VideoMLA replaces per-head KV caches with a shared low-rank content latent and decoupled 3D-RoPE key to cut memory 92.7 percent per token in video diffusion.

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

The paper introduces Multi-Head Latent Attention to video diffusion models to handle the high memory cost of maintaining KV caches during long autoregressive rollouts. It replaces standard per-head keys and values with one shared low-rank content latent plus a separate positional key that uses 3D rotary embeddings. This yields a 92.7 percent drop in per-token KV memory at every cached layer. The work further shows that the latent bottleneck itself enforces the effective rank during training, even though the pretrained attention weights are not low-rank, allowing quality to be retained at compression levels where simple spectral truncation would fail. Experiments on VBench confirm that short-horizon performance matches existing baselines while long-horizon results improve and generation speed increases by 1.23 times.

Core claim

VideoMLA replaces per-head keys and values with a shared low-rank content latent and a shared decoupled 3D-RoPE positional key. This reduces per-token KV memory by 92.7 percent at every cached layer. The method retains quality at compression ratios where direct spectral approximation would predict large reconstruction error because the MLA bottleneck, rather than the pretrained spectrum, determines the effective rank: both spectral and random initializations occupy nearly the full rank budget from initialization, and training preserves this budget while adapting within it.

What carries the argument

Multi-Head Latent Attention (MLA) that projects per-head keys and values into a shared low-rank content latent while handling positions with a separate decoupled 3D-RoPE key.

If this is right

  • Matches short-horizon streaming video diffusion baselines on VBench
  • Achieves the best overall score at long horizons among evaluated methods
  • Improves throughput by 1.23x on a single B200
  • Retains quality at high compression ratios even when pretrained attention is not low-rank

Where Pith is reading between the lines

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

  • The same latent compression pattern could be applied to autoregressive models in other modalities to achieve similar memory reductions without retraining from scratch.
  • Choosing the latent dimension may matter more than the choice of initialization for rank control in attention layers across diffusion architectures.
  • Minute-scale video generation may become feasible on hardware with tighter memory limits if the per-layer savings compound across deeper models.

Load-bearing premise

The low-rank latent bottleneck itself sets the effective rank during training rather than the spectrum of the pretrained attention weights.

What would settle it

Train or evaluate a video diffusion model with the same latent dimension but force a different rank occupation after training and check whether quality at the reported compression ratio collapses on long video sequences.

Figures

Figures reproduced from arXiv: 2605.30351 by Adil Kaan Akan, Hidir Yesiltepe, Hoda Eldardiry, Jiazhen Hu, Kaan Oktay, Pinar Yanardag, Tuna Han Salih Meral.

Figure 1
Figure 1. Figure 1: Pretrained video diffusion attention is not low-rank, unlike in language models. Singular value analysis of [WK; WV ] ∈ R 3072×1536 across the 30 transformer blocks of Wan2.1- T2V-1.3B. At dc = 192, the median layer captures only Emed = 0.458 of the spectral energy, and the 99%-energy effective rank exceeds 1300 in every layer. Self-Forcing [9] closed the train–test gap by conditioning training on self-gen… view at source ↗
Figure 2
Figure 2. Figure 2: The composed operator occupies its full rank-dc budget at every dc and every layer. Singular value analysis of the composed operator Mlearned = [W K ↑ W KV ↓ ; WV ↑ W KV ↓ ] for SVD￾initialized VideoMLA students at dc ∈ {64, 128, 256, 512}. (a) Median normalized spectra share a common envelope, truncated at dc. (b) Cumulative spectral energy. (c) Layer-wise 99%-energy effective rank: r0.99 ≈ 0.98 dc at eve… view at source ↗
Figure 3
Figure 3. Figure 3: Overview of VideoMLA. VideoMLA replaces dense per-head KV cache in Causal Wan 2.1- 1.3B with a low-rank latent obtained by jointly compressing keys and values through shared down/up projections, with positional information carried by a single decoupled rotated key. Orange blocks denote down projections, green blocks denote rotations, and white blocks denote up projections; latent frames are colored blue fo… view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative results. Samples generated by VideoMLA. Frames are shown at uniformly spaced timestamps from each 30s rollout, illustrating that the compressed latent KV cache preserves scene structure, subject identity, and visual fidelity over time. window followed by a weighted sum of vj,h produces the per-head output, and the head outputs are mixed through the output projection WO. The shape of score (h) i… view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative comparison. Long-rollout samples from VideoMLA and baseline causal video diffusion baselines under the same prompt. Each row shows uniformly spaced frames from one method. Model Results on 30s ↑ Results on 60s ↑ User Study ↑ AQ BC DD IQ MS SC Overall AQ BC DD IQ MS SC Overall PA TC DC Overall Self-Forcing [9] 0.541 0.948 0.624 0.577 0.952 0.932 0.762 0.565 0.958 0.393 0.650 0.987 0.974 0.755 2.… view at source ↗
Figure 6
Figure 6. Figure 6: Long-horizon generation quality. Frames sampled across a one-minute rollout of the same prompt. (Bottom) VideoMLA sustains visual fidelity with diverse, evolving motion, while (Top) LongSANA produces near-static content that degrades over time. VideoMLA yields higher visual fidelity and more diverse motion while achieving higher generation throughput and lower latency than LongSANA, and reduces KV cache si… view at source ↗
Figure 7
Figure 7. Figure 7: VideoMLA increases serving headroom under a fixed B200 memory budget. Compared with dense MHA, MLA greatly reduces per-batch memory growth and shifts the OOM limit to much larger batch sizes; the default dc = 192 gives 8.0× non-OOM batch headroom. 0 2 4 6 8 9.5 Training checkpoint step (k) 187 188 189 190 191 192 Mean effective rank r0.99 dc = 192 step 0 = init (a) Rank budget is saturated from initializat… view at source ↗
Figure 8
Figure 8. Figure 8: Rank-budget saturation during training. At dc = 192, both SVD and random initial￾ization occupy nearly the full latent rank budget from initialization, with stable effective rank and spectral tail throughout training. operator [WK; WV ] in Wan2.1-T2V-1.3B [PITH_FULL_IMAGE:figures/full_fig_p009_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: User Study Interface. User Study Interface for Long Video Generation Wan2.1-T2V-1.3B uses a diffusion transformer with multi-head self-attention over video latent tokens. In our implementation, the backbone contains 30 transformer blocks, hidden dimension d = 1536, nh = 12 attention heads, and per-head dimension dh = 128. In the dense baseline, each cached token stores both keys and values for all heads, g… view at source ↗
read the original abstract

Long-rollout causal video diffusion has converged on a fixed-size sliding-window KV cache, with recent progress innovating within this layout by changing which tokens occupy the window or how their positions are encoded. The per-head KV layout itself, a dominant contributor to streaming memory and latency, has been mostly left unchanged. In this paper, we present the first study of Multi-Head Latent Attention (MLA) in video diffusion. VideoMLA replaces per-head keys and values with a shared low-rank content latent and a shared decoupled 3D-RoPE positional key, reducing per-token KV memory by 92.7% at every cached layer. We further investigate why MLA succeeds in video diffusion even though the spectral assumption often used to motivate it in language models does not hold: pretrained video attention is not low-rank, with 99%-energy effective rank far above any practical latent dimension. VideoMLA retains quality at compression ratios where direct spectral approximation would predict large reconstruction error. We show that the MLA bottleneck, rather than the pretrained spectrum, determines the effective rank: both spectral and random initialization occupy nearly the full rank budget from initialization, and training preserves this budget while adapting within it. On VBench, VideoMLA matches short-horizon streaming video diffusion baselines, achieves the best overall score at long horizons among evaluated methods, and improves throughput by 1.23x on a single B200.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper introduces VideoMLA, an adaptation of Multi-Head Latent Attention (MLA) to autoregressive video diffusion. It replaces per-head keys and values with a shared low-rank content latent and a shared decoupled 3D-RoPE positional key, reducing per-token KV memory by 92.7% at every cached layer. Despite pretrained video attention having 99%-energy effective rank far above practical latent dimensions, VideoMLA retains quality on VBench at compression ratios where direct spectral approximation would predict large error. The authors attribute this to the MLA bottleneck determining effective rank, supported by the observation that both spectral and random initializations occupy nearly the full rank budget from the start and that training preserves this budget while adapting within it. Empirically, VideoMLA matches short-horizon streaming baselines, achieves the best overall VBench score at long horizons, and improves throughput by 1.23x on a single B200.

Significance. If the central claims hold, the work enables practical minute-scale autoregressive video diffusion by substantially lowering KV-cache memory and latency while preserving generation quality, with a concrete throughput gain. The empirical results on VBench for long-horizon settings and the investigation of rank dynamics constitute a useful contribution to efficient attention mechanisms in diffusion models. No machine-checked proofs or open reproducible code are mentioned.

major comments (2)
  1. [Abstract and Experiments] Abstract and Experiments: VBench scores and throughput numbers are reported without error bars, multiple runs, ablation studies on rank choice, or any description of training data and exclusion rules; the central quality-retention claim therefore rests on summary statistics whose robustness cannot be assessed.
  2. [Rank-dynamics analysis] Rank-dynamics analysis: The mechanistic claim that the MLA bottleneck (rather than the pretrained spectrum) determines effective rank rests on the specific observation that spectral and random initializations occupy nearly the full rank budget from initialization and that training preserves this budget; if this rank-occupation behavior does not generalize beyond the tested models or datasets, the attribution of quality retention to the bottleneck mechanism does not hold.
minor comments (1)
  1. [Abstract] Abstract: The statement that this is 'the first study of MLA in video diffusion' would benefit from a brief citation to prior MLA work in language models to clarify the novelty scope.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments. We address each major point below and indicate planned revisions where appropriate.

read point-by-point responses

  1. Referee: [Abstract and Experiments] Abstract and Experiments: VBench scores and throughput numbers are reported without error bars, multiple runs, ablation studies on rank choice, or any description of training data and exclusion rules; the central quality-retention claim therefore rests on summary statistics whose robustness cannot be assessed.

    Authors: We agree that the reported metrics lack error bars, multiple runs, rank ablations, and training-data details, limiting assessment of robustness. In revision we will rerun the main experiments across multiple random seeds and report means with standard deviations for both VBench scores and throughput. We will also add an ablation table varying latent rank and expand the experimental-setup section with a full description of the training corpus and exclusion criteria. revision: yes

  2. Referee: [Rank-dynamics analysis] Rank-dynamics analysis: The mechanistic claim that the MLA bottleneck (rather than the pretrained spectrum) determines effective rank rests on the specific observation that spectral and random initializations occupy nearly the full rank budget from initialization and that training preserves this budget; if this rank-occupation behavior does not generalize beyond the tested models or datasets, the attribution of quality retention to the bottleneck mechanism does not hold.

    Authors: The rank-occupation observations are reported for the concrete models and datasets used in the paper. We will revise the relevant section to explicitly limit the mechanistic attribution to the tested setting and to note that broader generalization remains an open question. The consistent behavior across both spectral and random initializations still supports the bottleneck explanation within the scope of our experiments. revision: partial

Circularity Check

0 steps flagged

No significant circularity; central claims rest on direct empirical measurements

full rationale

The paper's explanation for MLA quality retention at high compression is presented as an experimental observation: both spectral and random initializations occupy nearly the full rank budget from the start, with training preserving the budget while adapting inside it. This is a post-training measurement, not a fitted parameter renamed as a prediction or a self-definitional reduction. Performance results are reported via VBench scores and throughput benchmarks rather than any derivation that reduces to its own inputs by construction. No load-bearing self-citations, uniqueness theorems, or ansatzes smuggled via prior work appear in the derivation chain; the work is self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the architectural substitution and the empirical observation that the imposed low-rank bottleneck governs effective rank; no explicit free parameters, new axioms, or invented entities are named in the abstract.

axioms (1)
  • domain assumption Standard multi-head attention can be replaced by a low-rank latent without loss of expressivity when the bottleneck is the dominant constraint
    Invoked to explain why quality is retained despite high pretrained effective rank

pith-pipeline@v0.9.1-grok · 5814 in / 1405 out tokens · 22399 ms · 2026-06-29T08:15:19.926692+00:00 · methodology

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

31 extracted references · 24 canonical work pages · 12 internal anchors

  1. [1]

    In: Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing

    Ainslie, J., Lee-Thorp, J., De Jong, M., Zemlyanskiy, Y ., Lebrón, F., Sanghai, S.: Gqa: Training generalized multi-query transformer models from multi-head checkpoints. In: Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. pp. 4895–4901 (2023)

    2023

  2. [2]

    arXiv preprint arXiv:2602.10095 (2026)

    Bai, X., He, G., Li, Z., Shechtman, E., Huang, X., Wu, Z.: Causality in video diffusers is separable from denoising. arXiv preprint arXiv:2602.10095 (2026)

    work page arXiv 2026

  3. [3]

    Sana-video: Efficient video generation with block linear diffusion transformer, 2025

    Chen, J., Zhao, Y ., Yu, J., Chu, R., Chen, J., Yang, S., Wang, X., Pan, Y ., Zhou, D., Ling, H., et al.: Sana-video: Efficient video generation with block linear diffusion transformer. arXiv preprint arXiv:2509.24695 (2025)

    work page arXiv 2025

  4. [4]

    Self-Forcing++: Towards Minute-Scale High-Quality Video Generation

    Cui, J., Wu, J., Li, M., Yang, T., Li, X., Wang, R., Bai, A., Ban, Y ., Hsieh, C.J.: Self-forcing++: Towards minute-scale high-quality video generation. arXiv preprint arXiv:2510.02283 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  5. [5]

    Autoregressive Video Generation without Vector Quantization

    Deng, H., Pan, T., Diao, H., Luo, Z., Cui, Y ., Lu, H., Shan, S., Qi, Y ., Wang, X.: Autoregressive video generation without vector quantization. arXiv preprint arXiv:2412.14169 (2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  6. [6]

    arXiv preprint arXiv:2505.13544 (2025)

    Deng, K., Woodland, P.C.: Multi-head temporal latent attention. arXiv preprint arXiv:2505.13544 (2025)

    work page arXiv 2025

  7. [7]

    arXiv preprint arXiv:2508.03694 (2025)

    Gao, J., Chen, Z., Liu, X., Feng, J., Si, C., Fu, Y ., Qiao, Y ., Liu, Z.: Longvie: Multimodal-guided controllable ultra-long video generation. arXiv preprint arXiv:2508.03694 (2025)

    work page arXiv 2025

  8. [8]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Henschel, R., Khachatryan, L., Poghosyan, H., Hayrapetyan, D., Tadevosyan, V ., Wang, Z., Navasardyan, S., Shi, H.: Streamingt2v: Consistent, dynamic, and extendable long video gener- ation from text. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 2568–2577 (2025)

    2025

  9. [9]

    Self Forcing: Bridging the Train-Test Gap in Autoregressive Video Diffusion

    Huang, X., Li, Z., He, G., Zhou, M., Shechtman, E.: Self forcing: Bridging the train-test gap in autoregressive video diffusion. arXiv preprint arXiv:2506.08009 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  10. [10]

    In: Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (V olume 1: Long Papers)

    Ji, T., Guo, B., Wu, Y ., Guo, Q., Shen, L., Chen, Z., Qiu, X., Zhang, Q., Gui, T.: Towards economical inference: Enabling deepseek’s multi-head latent attention in any transformer- based llms. In: Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (V olume 1: Long Papers). pp. 33313–33328 (2025)

    2025

  11. [11]

    Pyramidal flow matching for efficient video generative modeling,

    Jin, Y ., Sun, Z., Li, N., Xu, K., Jiang, H., Zhuang, N., Huang, Q., Song, Y ., Mu, Y ., Lin, Z.: Pyra- midal flow matching for efficient video generative modeling. arXiv preprint arXiv:2410.05954 (2024)

    work page arXiv 2024

  12. [12]

    Memrope: Training-free infinite video generation via evolving memory tokens.arXiv preprint arXiv:2603.12513, 2026

    Kim, Y ., Hu, Q., Kuo, C.C.J., Beerel, P.A.: Memrope: Training-free infinite video generation via evolving memory tokens. arXiv preprint arXiv:2603.12513 (2026) 10

    work page arXiv 2026

  13. [13]

    Rolling Sink: Bridging Limited-Horizon Training and Open-Ended Testing in Autoregressive Video Diffusion

    Li, H., Liu, S., Lin, Z., Chandraker, M.: Rolling sink: Bridging limited-horizon training and open-ended testing in autoregressive video diffusion. arXiv preprint arXiv:2602.07775 (2026)

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  14. [14]

    DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model

    Liu, A., Feng, B., Wang, B., Wang, B., Liu, B., Zhao, C., Dengr, C., Ruan, C., Dai, D., Guo, D., et al.: Deepseek-v2: A strong, economical, and efficient mixture-of-experts language model. arXiv preprint arXiv:2405.04434 (2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  15. [15]

    DeepSeek-V3 Technical Report

    Liu, A., Feng, B., Xue, B., Wang, B., Wu, B., Lu, C., Zhao, C., Deng, C., Zhang, C., Ruan, C., et al.: Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437 (2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  16. [16]

    Rolling Forcing: Autoregressive Long Video Diffusion in Real Time

    Liu, K., Hu, W., Xu, J., Shan, Y ., Lu, S.: Rolling forcing: Autoregressive long video diffusion in real time. arXiv preprint arXiv:2509.25161 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  17. [17]

    Reward Forcing: Efficient Streaming Video Generation with Rewarded Distribution Matching Distillation

    Lu, Y ., Zeng, Y ., Li, H., Ouyang, H., Wang, Q., Cheng, K.L., Zhu, J., Cao, H., Zhang, Z., Zhu, X., et al.: Reward forcing: Efficient streaming video generation with rewarded distribution matching distillation. arXiv preprint arXiv:2512.04678 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  18. [18]

    arXiv preprint arXiv:2502.07864 (2025)

    Meng, F., Tang, P., Tang, X., Yao, Z., Sun, X., Zhang, M.: Transmla: Multi-head latent attention is all you need. arXiv preprint arXiv:2502.07864 (2025)

    work page arXiv 2025

  19. [19]

    OpenVid-1M: A Large-Scale High-Quality Dataset for Text-to-video Generation

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

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  20. [20]

    Advances in neural information processing systems30 (2017)

    Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, Ł., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems30 (2017)

    2017

  21. [21]

    Wan: Open and Advanced Large-Scale Video Generative Models

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

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  22. [22]

    LongLive: Real-time Interactive Long Video Generation

    Yang, S., Huang, W., Chu, R., Xiao, Y ., Zhao, Y ., Wang, X., Li, M., Xie, E., Chen, Y ., Lu, Y ., et al.: Longlive: Real-time interactive long video generation. arXiv preprint arXiv:2509.22622 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  23. [23]

    Anchor forcing: Anchor memory and tri-region rope for interactive streaming video diffusion

    Yang, Y ., Zhang, T., Huang, W., Chen, J., Wu, B., He, X., Cai, D., Li, B., Jiang, P.T.: Anchor forcing: Anchor memory and tri-region rope for interactive streaming video diffusion. arXiv preprint arXiv:2603.13405 (2026)

    work page arXiv 2026

  24. [24]

    Infinity-rope: Action-controllable infinite video generation emerges from autoregressive self- rollout.arXiv preprint arXiv:2511.20649, 2025

    Yesiltepe, H., Meral, T.H.S., Akan, A.K., Oktay, K., Yanardag, P.: Infinity-rope: Action- controllable infinite video generation emerges from autoregressive self-rollout. arXiv preprint arXiv:2511.20649 (2025)

    work page arXiv 2025

  25. [25]

    arXiv preprint arXiv:2512.05081 (2025) 4

    Yi, J., Jang, W., Cho, P.H., Nam, J., Yoon, H., Kim, S.: Deep forcing: Training-free long video generation with deep sink and participative compression. arXiv preprint arXiv:2512.05081 (2025)

    work page arXiv 2025

  26. [26]

    Advances in neural information processing systems37, 47455–47487 (2024)

    Yin, T., Gharbi, M., Park, T., Zhang, R., Shechtman, E., Durand, F., Freeman, B.: Improved distribution matching distillation for fast image synthesis. Advances in neural information processing systems37, 47455–47487 (2024)

    2024

  27. [27]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Yin, T., Zhang, Q., Zhang, R., Freeman, W.T., Durand, F., Shechtman, E., Huang, X.: From slow bidirectional to fast autoregressive video diffusion models. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 22963–22974 (2025)

    2025

  28. [28]

    arXiv preprint arXiv:2512.04519 , year=

    Yu, Y ., Wu, X., Hu, X., Hu, T., Sun, Y ., Lyu, X., Wang, B., Ma, L., Ma, Y ., Wang, Z., et al.: Videossm: Autoregressive long video generation with hybrid state-space memory. arXiv preprint arXiv:2512.04519 (2025)

    work page arXiv 2025

  29. [29]

    arXiv e-prints pp

    Zhang, L., Agrawala, M.: Packing input frame context in next-frame prediction models for video generation. arXiv e-prints pp. arXiv–2504 (2025)

    2025

  30. [30]

    Relax forcing: Relaxed kv-memory for consistent long video generation,

    Zhao, Z., Lu, Y ., Liu, Z., Song, J., Deng, J., Patras, I.: Relax forcing: Relaxed kv-memory for consistent long video generation. arXiv preprint arXiv:2603.21366 (2026) 11

    work page arXiv 2026

  31. [31]

    Causal Forcing: Autoregressive Diffusion Distillation Done Right for High-Quality Real-Time Interactive Video Generation

    Zhu, H., Zhao, M., He, G., Su, H., Li, C., Zhu, J.: Causal forcing: Autoregressive diffusion distillation done right for high-quality real-time interactive video generation. arXiv preprint arXiv:2602.02214 (2026) 12 Table of Contents A Videos and Website 1 B Details on User Study 1 C Background 1 C.1 Wan2.1-T2V-1.3B Backbone . . . . . . . . . . . . . . . ...

    work page internal anchor Pith review Pith/arXiv arXiv 2026