Pith. sign in

REVIEW 5 major objections 6 minor 98 references

LoViC: Efficient Long Video Generation with Context Compression

T0 review · 5 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read LoViC claims that a diffusion transformer can generate long, coherent videos by compressing the full history of prior segments into compact tokens, and demonstrates this for prediction, retrodiction, interpolation, and multi-shot…

desk verdict A genuinely useful compression module and a clean unified task design, but the paper's arbitrary-length, full-history claim outruns the evidence: every experiment uses at most two conditioning segments and never measures error accumulation beyond a third clip. read the letter →

arxiv 2507.12952 v1 pith:P4DTAZY6 submitted 2025-07-17 cs.CV

Jiaxiu Jiang , Wenbo Li , Jingjing Ren , Yuping Qiu , Yong Guo , Xiaogang Xu , Han Wu , Wangmeng Zuo This is my paper
classification cs.CV
keywords longvideogenerationdiffusiontransformercontextcompressionFlexFormerQ-Formerrotarypositionembeddingmulti-shottemporalcoherence
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

This paper tries to show that long-video generation does not have to pay the quadratic price of self-attention over the whole past. Its system, LoViC, generates video segment by segment, and before each new segment it compresses the entire preceding video and text history into a compact set of tokens with FlexFormer, a Q-Former-style autoencoder with an adjustable compression rate. The authors argue that this makes full-history conditioning practical, so a diffusion transformer can maintain long-range coherence at roughly constant per-step cost. They train on about 1.6 million open-domain clips and report that LoViC beats block-autoregressive baselines on continuity metrics while also supporting backward, interpolated, and multi-shot generation in one model. The central bet is that a lossy compressed memory retains enough spatial and semantic detail to keep identities and scenes stable across segments.

What carries the argument

FlexFormer is the component that carries the argument: a Q-Former-style autoencoder modified so that instead of cross-attention over a fixed query sequence, it uses self-attention with a single learnable query token replicated to match the input length. Paired with Interpolated-RoPE (I-RoPE), which gives each compressed query a spatiotemporal position by interpolating the positions of video tokens, this yields an adjustable compression ratio that can be uniform, linear, or log across time. These compressed tokens are injected into the DiT's self-attention, so the generator sees the whole history at a cost set by the compression ratio rather than by the length of the history. The linear compression strategy, which keeps more tokens near the current segment, is what the paper ultimately adopts for its main results.

What would settle it

Compress a video in which a distinctive object appears only in the early context at several compression ratios, generate the next segment with the DiT held fixed, and count how often the object's identity survives; if the survival rate collapses as compression gets lossier, the sufficiency of compressed context for long-range coherence is falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that a DiT can be conditioned on arbitrarily long video-text context without quadratic cost, if that context is first compressed by a dedicated autoencoder. At each generation step, FlexFormer encodes the history of preceding segments, together with their text prompts, into context tokens; a single learnable query token is replicated according to the input length, and Interpolated-RoPE assigns each query a spatiotemporal position interpolated from the video tokens so spatial structure survives compression. The context tokens are concatenated into the self-attention of every DiT block, avoiding extra cross-attention layers, and a temporal gap in the positional encoding separates shots for multi-shot generation. The paper demonstrates this unified paradigm on video prediction, retrodiction, interpolation, and multi-shot generation, and argues that its continuity gains come precisely from being able to use the full history rather than a short conditioning window.

Load-bearing premise

The load-bearing premise is that the compressed context tokens preserve enough spatial, temporal, and semantic detail about the history; if the compression silently drops an object's identity or the scene layout, the next segment cannot be coherent no matter how cheap the attention becomes.

Editorial extensions

If this is right

  • Longer conditioning context should improve temporal consistency, as the paper observes that models with larger context windows score higher on subject and background consistency.
  • Compressing history does not necessarily weaken generation: LoViC outperforms a baseline that consumes the full uncompressed conditioning video, which the paper reads as evidence that conditioning features contain redundancy.
  • One trained model can switch between prediction, retrodiction, interpolation, and multi-shot generation by changing only the positional encoding and the sampled training task.
  • The segment-wise design should extend a fixed-length base model to arbitrarily many frames, with memory and runtime per step held roughly constant by the compression.

Reading between the lines

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

  • A testable extension the paper does not run would degrade FlexFormer's compression fidelity in a controlled way and measure whether downstream continuity falls correspondingly, isolating how much of the gain is due to compression quality rather than the DiT's own inductive biases.
  • Because text is compressed through the same pathway as video, prompt adherence may weaken as context grows; the paper reports lower video-text alignment than its base model in multi-shot generation, consistent with this pressure.
  • The fixed linear compression schedule suggests a learnable successor: a model could decide which regions of history will need detail for the next segment, rather than allocating tokens by a hand-set rule.
  • FlexFormer can also be read as a memory mechanism for video world models, with the compressed context playing the role of a recurrent state updated every segment.
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

5 major / 6 minor

Summary. The paper introduces LoViC, a diffusion-transformer framework for long-video generation that generates videos segment by segment while conditioning on previously generated segments through compressed context tokens. The core component, FlexFormer, is a Q-Former-style autoencoder with a single learnable query token and an interpolated RoPE (I-RoPE) positional encoding, designed to compress variable-length video-text contexts at adjustable compression ratios. LoViC is trained on a re-captioned Panda-70M subset and evaluated on video prediction, retrodiction, interpolation, and multi-shot generation. The paper reports favorable quantitative results on 100 test videos per task, along with ablations of FlexFormer components and compression strategies.

Significance. If the central claims are validated, LoViC would be a practically useful way to extend DiT-based video generation to longer horizons: compressed context tokens injected into self-attention avoid extra cross-attention parameters, and the unified treatment of prediction, retrodiction, interpolation, and multi-shot generation is attractive. The paper also has concrete strengths: training on a million-scale open-domain corpus with re-captioned descriptions, a clear three-stage training recipe, and honest discussion of limitations, including reconstruction loss from the base VAE. However, the headline capability of conditioning on the full history of preceding segments at arbitrary length is not actually demonstrated by the experiments, which use at most two conditioning segments and a single level of autoregressive rollout. The empirical evaluation is also narrow, with no confidence intervals or significance tests, and the multi-shot baselines do not support video conditioning, making the comparison asymmetric. These gaps prevent the paper from fully supporting its central claims.

major comments (5)
  1. [Introduction / Context Compression with FlexFormer / Evaluation Settings] The claim that the model 'conditions on the full history of preceding video segments' and compresses 'context of arbitrary length' is not supported by the experiments. In every evaluated task the conditioning history is at most two segments: prediction uses the first two clips, interpolation uses the first and last clips, retrodiction uses the last two clips, and multi-shot generation produces only a second and third clip from the first clip. The method section states that each (Video_i, Text_i) pair is encoded independently and the compressed tokens are concatenated, so the number of context tokens grows linearly with the number of segments unless a global compression over accumulated history is introduced; no such policy is described or tested. The paper's own conclusion defers 'architecture that can adaptively extract and integrate relevant information from previously generated video segments' to future work. The full-history, arbitrary-length property is therefore an extrapolation from a three-clip, mostly teacher-forced regime, and long-horizon error accumulation from conditioning on self-generated segments is unmeasured. Either add experiments with more than two conditioning segments and longer autoregressive rollouts, or reframe the central claim to the demonstrated regime.
  2. [Context Compression with FlexFormer / Table 3 / Table 4] The load-bearing premise that FlexFormer's compressed tokens preserve enough spatial, temporal, and semantic detail for the downstream DiT is only indirectly validated. Table 3 shows real reconstruction loss (PSNR 21.01, LPIPS 0.244, with the caveat that the base VAE contributes), yet no experiment isolates how generation quality varies with compression fidelity while holding the conditioning history fixed. Table 4 varies compression strategies and ratios, but it does so in the same three-clip setting and mixes changes in strategy with changes in ratio. The paper would be substantially strengthened by a controlled comparison of uncompressed conditioning versus compressed conditioning, or by varying the number of conditioning segments while measuring both cost and generation metrics. Without such an experiment, the claim that context compression preserves the information needed for long-range consistency is not directly established.
  3. [Evaluation Settings / Tables 1 and 2] All quantitative comparisons are based on 100 test videos per task and are reported without error bars, confidence intervals, or significance tests. Several differences in Table 1 are small (e.g., PSNR 15.76 vs. 15.70, Video-text Alignment 0.237 vs. 0.220, LPIPS 0.316 vs. 0.344 in the prediction column), so the statement that the model 'outperforms all baselines' on continuity metrics is not statistically supported. Table 2 shows that the proposed method is actually lower than LTX-Video on Aesthetic Quality (0.448 vs. 0.483) and Video-text Alignment (0.180 vs. 0.231), which the text acknowledges, but the overall narrative would benefit from explicit variance reporting and tests for the claimed improvements.
  4. [Multi-Shot Video Generation / Table 2] The multi-shot comparison is asymmetric: both baselines in Table 2, MEVG and LTX-Video, 'do not support video-conditioned generation' as the caption states. Therefore the higher Subject Consistency and Background Consistency of LoViC could be due to the presence of video conditioning alone, rather than to FlexFormer compression or the proposed architecture. To isolate the contribution of the compression mechanism, the authors should compare against a video-conditioned baseline without compression, or include an ablated variant of LoViC that consumes uncompressed context under the same experimental setup. As it stands, the multi-shot results demonstrate the benefit of video conditioning but do not specifically validate the FlexFormer contribution.
  5. [Implementation Details / Ablation Study] Some key hyperparameters appear to be selected using the same evaluation metrics that are later reported as results. The temporal gap of 20 latent frames for multi-shot generation is described as 'empirically chosen,' and the linear compression strategy is adopted 'for the results reported in Table 1' after comparing strategies in Table 4. The paper does not state whether these choices were made on a held-out validation set or on the same 100-video test sets used for the final numbers. If selection was performed on the test sets, the reported metrics are optimistically biased. Please specify the validation procedure and, where possible, report results for the non-selected settings on a separate test split.
minor comments (6)
  1. [Abstract and Introduction] The word 'retradiction' appears in the Abstract and Introduction and should be 'retrodiction'; 'conditon' appears in Evaluation Settings and should be 'condition'.
  2. [Equation (3)] The RoPE frequency formula appears to show an incorrect exponent sign: standard RoPE uses theta_i = 10000^{-2i/d}, not 10000^{2i/d}. Please correct the equation and the surrounding notation.
  3. [Figure 5] The caption and text refer to left and middle panels; it would be clearer to label all three panels (left, middle, right) explicitly, since Figure 5 appears to show three compression strategies.
  4. [Experimental Setup / Dataset Construction] The sentence 'Panda-70M ... which builds the ASR-captioned video-text dataset HD-VILA-100M' is grammatically unclear; it should say 'which builds on' or similar.
  5. [Table 1] The caption explains the asterisk for unfixed context length and the dash for unsupported settings, but the row labels could also state which tasks each baseline supports; currently the reader must infer this from the dash patterns.
  6. [Figure 2] The memory/time comparison in Figure 2 would be more informative if the hardware, batch size, and sequence lengths used for the profiling were specified in the caption.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper is an empirical architecture study whose FlexFormer compression and DiT generation are trained and evaluated independently of the claims being made.

full rationale

LoViC's derivation chain is self-contained. FlexFormer is trained from scratch with MSE loss in stage 1; the DiT is trained with the flow-matching loss in Eqs. (1)-(2) in stages 2-3; and the evaluation compares generated clips against ground-truth clips (PSNR/SSIM/LPIPS) and VBench metrics. There is no equation whose target quantity is defined by the fitted parameters, and no fitted parameter is renamed as a prediction. The only author self-citation (Guo et al. 2025, sharing co-author Yong Guo) appears in two places: 'Similar positional strategies have also been adopted in multi-shot video generation' and a general list of prior work showing context matters; neither carries the architecture's correctness or the experimental results, so it is not load-bearing. The paper's real weakness is a scope mismatch, not circularity: the 'full history ... arbitrary length' claim is evaluated only on three clips (two conditioning segments), FlexFormer encodes each (Video_i, Text_i) pair independently and concatenates context tokens, and the conclusion defers 'architecture that can adaptively extract and integrate relevant information from previously generated video segments' to future work. That under-tests the headline capability, but it does not make any result equivalent to its inputs by construction. The choice of the linear compression strategy via Table 4 and the 20-latent-frame temporal gap are selected on evaluation metrics, which is a test-set-selection concern rather than a circular reduction.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

No new physical entities are introduced. FlexFormer and I-RoPE are architectural components, not invented particles, forces, or dimensions. The load-bearing free parameters are the compression schedule and the multi-shot temporal gap, both chosen empirically rather than derived.

free parameters (3)
  • Temporal gap for multi-shot = 20 latent frames
    Empirically chosen in the third training stage to separate shots in positional space; not derived from a principled scale.
  • Linear compression schedule ratios = 5.4 (16->1), 3.4 (8->1), and log variants 9.7 and 5.8 in Table 4
    The compression rate and distribution of query tokens along time are chosen by hand or ablation; the linear strategy is selected because it performs best in Table 4, a post-hoc selection on the test set.
  • ViCLIP cosine similarity threshold = 0.5
    Used to filter video triplets during dataset construction; roughly half of the data is discarded, potentially biasing the training distribution.
assumptions (4)
  • standard math Flow matching objective (Eq. 1-2) is a valid training signal for the DiT with compressed context features.
    Standard flow-matching formulation used in modern video DiTs; accepted background.
  • domain assumption A single learnable query token, replicated and position-encoded, can summarize multi-modal video-text context with sufficient fidelity.
    Core to FlexFormer; only validated empirically via reconstruction PSNR/LPIPS and downstream generation metrics on a small test set.
  • domain assumption I-RoPE interpolated positions capture the spatial structure needed for compression, while treating query tokens as 1D text tokens (M-RoPE) is insufficient.
    Supported only by Table 3 ablation; the paper offers a hypothesis, not a proof, for why M-RoPE underperforms.
  • domain assumption LTX-Video V0.9.0 VAE compression quality is adequate for the pipeline.
    The FlexFormer reconstruction is cascaded with the base VAE; the paper notes reconstruction performance is affected by the VAE (Table 3 caption).

how reviews work

0 comments
Cite this review

Pith. "Pith review of LoViC: Efficient Long Video Generation with Context Compression." pith.science (2026). https://pith.science/paper/P4DTAZY6

@misc{pith2026250712952,
  author       = {Pith},
  title        = {Pith review of: LoViC: Efficient Long Video Generation with Context Compression},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/P4DTAZY6}},
  note         = {Machine review of arXiv:2507.12952}
}
read the original abstract

Despite recent advances in diffusion transformers (DiTs) for text-to-video generation, scaling to long-duration content remains challenging due to the quadratic complexity of self-attention. While prior efforts -- such as sparse attention and temporally autoregressive models -- offer partial relief, they often compromise temporal coherence or scalability. We introduce LoViC, a DiT-based framework trained on million-scale open-domain videos, designed to produce long, coherent videos through a segment-wise generation process. At the core of our approach is FlexFormer, an expressive autoencoder that jointly compresses video and text into unified latent representations. It supports variable-length inputs with linearly adjustable compression rates, enabled by a single query token design based on the Q-Former architecture. Additionally, by encoding temporal context through position-aware mechanisms, our model seamlessly supports prediction, retradiction, interpolation, and multi-shot generation within a unified paradigm. Extensive experiments across diverse tasks validate the effectiveness and versatility of our approach.

Figures

Figures reproduced from arXiv: 2507.12952 by the authors.

Figure 1
Figure 1. Videos generated by our model. Our model has the flexibility to do video continuation of any direction and generate multi-shot video with advanced efficiency. It shows the capability of retaining ID consistency within large temporal range, generating videos of large and smooth motion. Abstract Despite recent advances in diffusion transformers (DiTs) for text-to-video generation, scaling to long-duration content re￾m… view at source ↗
Figure 2
Figure 2. Memory and time usage of single timestep DiT inference. With context compression, our method reduces memory usage and runtime, allowing more frames to be gen￾erated within the same resource constraints. first-person driving simulations, immersive 3D games, prod￾uct promotional content, and user-generated storytelling— demand coherent long-form video generation. Naively ex￾tending DiT models to longer sequences is co… view at source ↗
Figure 3
Figure 3. Model architecture. The left part of the figure features an autoencoder consisting of a FlexFormer encoder and a FlexFormer decoder. The encoder compresses multiple segments of video and text tokens separately. The number of query tokens is derived from the video token sequence length. The query token sequence is formed by copying the single learnable token multiple times. The decoder decode the context tokens into … view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Illustration of our positional encoding. Each block represents the positional index (t, h, w) of the corresponding token. The illustrated compression strategy is uniform compression. To adapt for multi-shot generation, the blue and purple blocks will be separated sligh…
Figure 5
Figure 5. Figure 5: Illustration of different compression strategies. Blue and purple dots represent the position of video tokens and query tokens respectively. Text tokens are omitted. in [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: Comparison on video prediction, interpolation and retrodiction. Our model demonstrates the best consis￾tency between conditioning video and generated video and produces a smooth transition between the videos. among various generation tasks. For multi-shot video gener￾a…
Figure 7
Figure 7. Figure 7: Comparison on multi-shot video generation. The baselines do not support conditioning on video. 0.5, which are approximately half the dataset. For video prediction, interpolation and retrodiction tasks, we further segment each single-shot video in the 2M sub￾set into th…
Figure 8
Figure 8. Figure 8 [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]
Figure 9
Figure 9. Figure 9: Ablation of FlexFormer components. Flex￾Former is cascaded with VAE to reconstruct the videos. M-RoPE to differentiate the positions of video features and learnable query tokens, where the queries are treated as text tokens. When replacing the multiple query tokens wit…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

98 extracted references · 12 canonical work pages

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    S.; and Vanden-Eijnden, E

    Albergo, M. S.; and Vanden-Eijnden, E. 2022. Building normalizing flows with stochastic interpolants. arXiv preprint arXiv:2209.15571

    arXiv 2022

  4. [4]

    Bai, S.; Chen, K.; Liu, X.; Wang, J.; Ge, W.; Song, S.; Dang, K.; Wang, P.; Wang, S.; Tang, J.; et al. 2025. Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923

    arXiv 2025

  5. [5]

    Bansal, H.; Bitton, Y.; Yarom, M.; Szpektor, I.; Grover, A.; and Chang, K.-W. 2024. Talc: Time-aligned captions for multi-scene text-to-video generation. arXiv preprint arXiv:2405.04682

    arXiv 2024

  6. [6]

    Bao, F.; Nie, S.; Xue, K.; Cao, Y.; Li, C.; Su, H.; and Zhu, J. 2023. All are worth words: A vit backbone for diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 22669--22679

    2023

  7. [7]

    Bar-Tal, O.; Chefer, H.; Tov, O.; Herrmann, C.; Paiss, R.; Zada, S.; Ephrat, A.; Hur, J.; Liu, G.; Raj, A.; et al. 2024. Lumiere: A space-time diffusion model for video generation. In SIGGRAPH Asia 2024 Conference Papers, 1--11

    2024

  8. [8]

    Blattmann, A.; Dockhorn, T.; Kulal, S.; Mendelevitch, D.; Kilian, M.; Lorenz, D.; Levi, Y.; English, Z.; Voleti, V.; Letts, A.; et al. 2023 a . Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127

    arXiv 2023

Show all 98 references
  1. [9]

    W.; Fidler, S.; and Kreis, K

    Blattmann, A.; Rombach, R.; Ling, H.; Dockhorn, T.; Kim, S. W.; Fidler, S.; and Kreis, K. 2023 b . Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 22563--22575

    2023

  2. [10]

    Chen, B.; Mart \' Mons \'o , D.; Du, Y.; Simchowitz, M.; Tedrake, R.; and Sitzmann, V. 2024 a . Diffusion forcing: Next-token prediction meets full-sequence diffusion. Advances in Neural Information Processing Systems, 37: 24081--24125

    2024

  3. [11]

    Chen, G.; Lin, D.; Yang, J.; Lin, C.; Zhu, J.; Fan, M.; Zhang, H.; Chen, S.; Chen, Z.; Ma, C.; et al. 2025 a . Skyreels-v2: Infinite-length film generative model. arXiv preprint arXiv:2504.13074

    2025 arXiv

  4. [12]

    Chen, H.; Xia, M.; He, Y.; Zhang, Y.; Cun, X.; Yang, S.; Xing, J.; Liu, Y.; Chen, Q.; Wang, X.; et al. 2023 a . Videocrafter1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512

    2023 arXiv

  5. [13]

    Chen, H.; Zhang, Y.; Cun, X.; Xia, M.; Wang, X.; Weng, C.; and Shan, Y. 2024 b . Videocrafter2: Overcoming data limitations for high-quality video diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 7310--7320

    2024

  6. [14]

    Chen, J.; Long, F.; An, J.; Qiu, Z.; Yao, T.; Luo, J.; and Mei, T. 2025 b . Ouroboros-Diffusion: Exploring Consistent Content Generation in Tuning-free Long Video Diffusion. arXiv preprint arXiv:2501.09019

    2025 arXiv

  7. [15]

    E.; Fang, Y.; Lee, H.-Y.; Ren, J.; Yang, M.-H.; et al

    Chen, T.-S.; Siarohin, A.; Menapace, W.; Deyneka, E.; Chao, H.-w.; Jeon, B. E.; Fang, Y.; Lee, H.-Y.; Ren, J.; Yang, M.-H.; et al. 2024 c . Panda-70m: Captioning 70m videos with multiple cross-modality teachers. In Proceedings of the IEEE/CVF Conference on Computer Vision and ...

    2024

  8. [16]

    Chen, X.; Wang, Y.; Zhang, L.; Zhuang, S.; Ma, X.; Yu, J.; Wang, Y.; Lin, D.; Qiao, Y.; and Liu, Z. 2023 b . Seine: Short-to-long video diffusion model for generative transition and prediction. In The Twelfth International Conference on Learning Representations

    2023

  9. [17]

    C.; Kautz, J.; Guestrin, C.; et al

    Dalal, K.; Koceja, D.; Hussein, G.; Xu, J.; Zhao, Y.; Song, Y.; Han, S.; Cheung, K. C.; Kautz, J.; Guestrin, C.; et al. 2025. One-minute video generation with test-time training. arXiv preprint arXiv:2504.05298

    2025 arXiv

  10. [18]

    Esser, P.; Kulal, S.; Blattmann, A.; Entezari, R.; M \"u ller, J.; Saini, H.; Levi, Y.; Lorenz, D.; Sauer, A.; Boesel, F.; et al. 2024. Scaling rectified flow transformers for high-resolution image synthesis. In Forty-first international conference on machine learning

    2024

  11. [19]

    Fang, X.; Ma, L.; Chen, Z.; Zhou, M.; and Qi, G.-j. 2025. InfLVG: Reinforce Inference-Time Consistent Long Video Generation with GRPO. arXiv preprint arXiv:2505.17574

    2025 arXiv

  12. [20]

    Gao, K.; Shi, J.; Zhang, H.; Wang, C.; and Xiao, J. 2024 a . Vid-gpt: Introducing gpt-style autoregressive generation in video diffusion models. arXiv preprint arXiv:2406.10981

    2024 arXiv

  13. [21]

    Gao, K.; Shi, J.; Zhang, H.; Wang, C.; Xiao, J.; and Chen, L. 2024 b . Ca2-VDM: Efficient Autoregressive Video Diffusion Model with Causal Generation and Cache Sharing. arXiv preprint arXiv:2411.16375

    2024 arXiv

  14. [22]

    Gu, J.; Wang, S.; Zhao, H.; Lu, T.; Zhang, X.; Wu, Z.; Xu, S.; Zhang, W.; Jiang, Y.-G.; and Xu, H. 2023. Reuse and diffuse: Iterative denoising for text-to-video generation. arXiv preprint arXiv:2309.03549

    2023 arXiv

  15. [23]

    Gu, Y.; Mao, W.; and Shou, M. Z. 2025. Long-context autoregressive video modeling with next-frame prediction. arXiv preprint arXiv:2503.19325

    2025 arXiv

  16. [24]

    Guo, Y.; Yang, C.; Rao, A.; Liang, Z.; Wang, Y.; Qiao, Y.; Agrawala, M.; Lin, D.; and Dai, B. 2023. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725

    2023 arXiv

  17. [25]

    Guo, Y.; Yang, C.; Yang, Z.; Ma, Z.; Lin, Z.; Yang, Z.; Lin, D.; and Jiang, L. 2025. Long context tuning for video generation. arXiv preprint arXiv:2503.10589

    2025 arXiv

  18. [26]

    HaCohen, Y.; Chiprut, N.; Brazowski, B.; Shalem, D.; Moshe, D.; Richardson, E.; Levin, E.; Shiran, G.; Zabari, N.; Gordon, O.; et al. 2024. Ltx-video: Realtime video latent diffusion. arXiv preprint arXiv:2501.00103

    2024 arXiv

  19. [27]

    Harvey, W.; Naderiparizi, S.; Masrani, V.; Weilbach, C.; and Wood, F. 2022. Flexible diffusion modeling of long videos. Advances in Neural Information Processing Systems, 35: 27953--27965

    2022

  20. [28]

    Henschel, R.; Khachatryan, L.; Hayrapetyan, D.; Poghosyan, H.; Tadevosyan, V.; Wang, Z.; Navasardyan, S.; and Shi, H. 2024. Streamingt2v: Consistent, dynamic, and extendable long video generation from text. arXiv preprint arXiv:2403.14773

    2024 arXiv

  21. [29]

    P.; Poole, B.; Norouzi, M.; Fleet, D

    Ho, J.; Chan, W.; Saharia, C.; Whang, J.; Gao, R.; Gritsenko, A.; Kingma, D. P.; Poole, B.; Norouzi, M.; Fleet, D. J.; and Salimans, T. 2022 a . Imagen Video: High Definition Video Generation with Diffusion Models. arXiv preprint arXiv: 2210.02303

    2022 arXiv

  22. [30]

    Ho, J.; Salimans, T.; Gritsenko, A.; Chan, W.; Norouzi, M.; and Fleet, D. J. 2022 b . Video diffusion models. Advances in Neural Information Processing Systems, 35: 8633--8646

    2022

  23. [31]

    Hu, J.; Hu, S.; Song, Y.; Huang, Y.; Wang, M.; Zhou, H.; Liu, Z.; Ma, W.-Y.; and Sun, M. 2024. ACDiT: Interpolating Autoregressive Conditional Modeling and Diffusion Transformer. arXiv preprint arXiv:2412.07720

    2024

  24. [32]

    Huang, X.; Li, Z.; He, G.; Zhou, M.; and Shechtman, E. 2025. Self Forcing: Bridging the Train-Test Gap in Autoregressive Video Diffusion. arXiv preprint arXiv:2506.08009

    2025 arXiv

  25. [33]

    Huang, Z.; He, Y.; Yu, J.; Zhang, F.; Si, C.; Jiang, Y.; Zhang, Y.; Wu, T.; Jin, Q.; Chanpaisit, N.; et al. 2024. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 21807--21818

    2024

  26. [34]

    Jin, Y.; Sun, Z.; Li, N.; Xu, K.; Jiang, H.; Zhuang, N.; Huang, Q.; Song, Y.; Mu, Y.; and Lin, Z. 2024. Pyramidal flow matching for efficient video generative modeling. arXiv preprint arXiv:2410.05954

    2024

  27. [35]

    Kang, T.; Kothandaraman, D.; and Lin, M. C. 2025. Text2story: Advancing video storytelling with text guidance. arXiv preprint arXiv:2503.06310

    2025 arXiv

  28. [36]

    K.; Liu, F.; Ceylan, D.; Rehg, J

    Kara, O.; Singh, K. K.; Liu, F.; Ceylan, D.; Rehg, J. M.; and Hinz, T. 2025. ShotAdapter: Text-to-Multi-Shot Video Generation with Diffusion Models. arXiv preprint arXiv:2505.07652

    2025 arXiv

  29. [37]

    Kim, J.; Kang, J.; Choi, J.; and Han, B. 2025 a . FIFO-Diffusion: Generating Infinite Videos from Text without Training. Advances in Neural Information Processing Systems, 37: 89834--89868

    2025

  30. [38]

    W.; Wang, J.-H.; Lee, J.-Y.; and Shin, J

    Kim, S.; Oh, S. W.; Wang, J.-H.; Lee, J.-Y.; and Shin, J. 2025 b . Tuning-Free Multi-Event Long Video Generation via Synchronized Coupled Sampling. arXiv preprint arXiv:2503.08605

    2025 arXiv

  31. [39]

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

    2024 arXiv

  32. [40]

    Li, J.; Li, D.; Savarese, S.; and Hoi, S. 2023. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, 19730--19742. PMLR

    2023

  33. [41]

    Li, Y.; Beluch, W.; Keuper, M.; Zhang, D.; and Khoreva, A. 2024 a . Vstar: Generative temporal nursing for longer dynamic video synthesis. arXiv preprint arXiv:2403.13501

    2024 arXiv

  34. [42]

    Li, Z.; Hu, S.; Liu, S.; Zhou, L.; Choi, J.; Meng, L.; Guo, X.; Li, J.; Ling, H.; and Wei, F. 2024 b . Arlon: Boosting diffusion transformers with autoregressive models for long video generation. arXiv preprint arXiv:2410.20502

    2024 arXiv

  35. [43]

    Lin, B.; Ge, Y.; Cheng, X.; Li, Z.; Zhu, B.; Wang, S.; He, X.; Ye, Y.; Yuan, S.; Chen, L.; et al. 2024. Open-sora plan: Open-source large video generation model. arXiv preprint arXiv:2412.00131

    2024 arXiv

  36. [44]

    Lin, S.; Yang, C.; He, H.; Jiang, J.; Ren, Y.; Xia, X.; Zhao, Y.; Xiao, X.; and Jiang, L. 2025. Autoregressive Adversarial Post-Training for Real-Time Interactive Video Generation. arXiv preprint arXiv:2506.09350

    2025

  37. [45]

    T.; Ben-Hamu, H.; Nickel, M.; and Le, M

    Lipman, Y.; Chen, R. T.; Ben-Hamu, H.; Nickel, M.; and Le, M. 2022. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747

    2022 arXiv

  38. [46]

    Liu, D.; Li, S.; Liu, Y.; Li, Z.; Wang, K.; Li, X.; Qin, Q.; Liu, Y.; Xin, Y.; Li, Z.; et al. 2025. Lumina-Video: Efficient and Flexible Video Generation with Multi-scale Next-DiT. arXiv preprint arXiv:2502.06782

    2025 arXiv

  39. [47]

    Liu, X.; Gong, C.; and Liu, Q. 2022. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003

    2022 arXiv

  40. [48]

    Long, F.; Qiu, Z.; Yao, T.; and Mei, T. 2024. VideoStudio: Generating Consistent-Content and Multi-Scene Videos. In European Conference on Computer Vision, 468--485. Springer

    2024

  41. [49]

    ???? FreeLong: Training-Free Long Video Generation with SpectralBlend Temporal Attention

    Lu, Y.; Liang, Y.; Zhu, L.; and Yang, Y. ???? FreeLong: Training-Free Long Video Generation with SpectralBlend Temporal Attention. In The Thirty-eighth Annual Conference on Neural Information Processing Systems

  42. [50]

    Ma, G.; Huang, H.; Yan, K.; Chen, L.; Duan, N.; Yin, S.; Wan, C.; Ming, R.; Song, X.; Chen, X.; et al. 2025. Step-video-t2v technical report: The practice, challenges, and future of video foundation model. arXiv preprint arXiv:2502.10248

    2025 arXiv

  43. [51]

    Oh, G.; Jeong, J.; Kim, S.; Byeon, W.; Kim, J.; Kim, S.; and Kim, S. 2024. Mevg: Multi-event video generation with text-to-video models. In European Conference on Computer Vision, 401--418. Springer

    2024

  44. [52]

    Ouyang, Y.; Zhao, H.; Wang, G.; et al. 2024. Flexifilm: Long video generation with flexible conditions. arXiv preprint arXiv:2404.18620

    2024 arXiv

  45. [53]

    Peebles, W.; and Xie, S. 2023. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF international conference on computer vision, 4195--4205

    2023

  46. [54]

    Polyak, A.; Zohar, A.; Brown, A.; Tjandra, A.; Sinha, A.; Lee, A.; Vyas, A.; Shi, B.; Ma, C.-Y.; Chuang, C.-Y.; et al. 2024. Movie Gen: A Cast of Media Foundation Models. arXiv preprint arXiv:2410.13720

    2024 arXiv

  47. [55]

    Qi, T.; Yuan, J.; Feng, W.; Fang, S.; Liu, J.; Zhou, S.; He, Q.; Xie, H.; and Zhang, Y. 2025. Mask2DiT: Dual Mask-based Diffusion Transformer for Multi-Scene Long Video Generation. arXiv preprint arXiv:2503.19881

    2025 arXiv

  48. [56]

    Qiu, H.; Xia, M.; Zhang, Y.; He, Y.; Wang, X.; Shan, Y.; and Liu, Z. 2023. Freenoise: Tuning-free longer video diffusion via noise rescheduling. arXiv preprint arXiv:2310.15169

    2023 arXiv

  49. [57]

    Rombach, R.; Blattmann, A.; Lorenz, D.; Esser, P.; and Ommer, B. 2022. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 10684--10695

    2022

  50. [58]

    Ronneberger, O.; Fischer, P.; and Brox, T. 2015. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention--MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part II...

    2015

  51. [59]

    ???? Rolling Diffusion Models

    Ruhe, D.; Heek, J.; Salimans, T.; and Hoogeboom, E. ???? Rolling Diffusion Models. In Forty-first International Conference on Machine Learning

  52. [60]

    P.; Wang, X.; and Gool, L

    Savov, N.; Kazemi, N.; Zhang, D.; Paudel, D. P.; Wang, X.; and Gool, L. V. 2025. StateSpaceDiffuser: Bringing Long Context to Diffusion World Models. arXiv preprint arXiv: 2505.22246

    2025

  53. [61]

    Song, J.; Meng, C.; and Ermon, S. 2020. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502

    2020 arXiv

  54. [62]

    Song, K.; Chen, B.; Simchowitz, M.; Du, Y.; Tedrake, R.; and Sitzmann, V. 2025. History-Guided Video Diffusion. arXiv preprint arXiv:2502.06764

    2025 arXiv

  55. [63]

    Su, J.; Ahmed, M.; Lu, Y.; Pan, S.; Bo, W.; and Liu, Y. 2024. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568: 127063

    2024

  56. [64]

    J.; Li, Y.; Jia, X.; Adam, H.; Hariharan, B.; Zhao, L.; and Liu, T

    Sun, Y.; Zhou, H.; Yuan, L.; Sun, J. J.; Li, Y.; Jia, X.; Adam, H.; Hariharan, B.; Zhao, L.; and Liu, T. 2024. Video Creation by Demonstration. arXiv preprint arXiv:2412.09551

    2024 arXiv

  57. [65]

    Teng, H.; Jia, H.; Sun, L.; Li, L.; Li, M.; Tang, M.; Han, S.; Zhang, T.; Zhang, W.; Luo, W.; et al. 2025. MAGI-1: Autoregressive Video Generation at Scale. arXiv preprint arXiv:2505.13211

    2025 arXiv

  58. [66]

    Voleti, V.; Jolicoeur-Martineau, A.; and Pal, C. 2022. Mcvd-masked conditional video diffusion for prediction, generation, and interpolation. Advances in neural information processing systems, 35: 23371--23385

    2022

  59. [67]

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

    2025 arXiv

  60. [68]

    Wang, F.-Y.; Chen, W.; Song, G.; Ye, H.-J.; Liu, Y.; and Li, H. 2023 a . Gen-l-video: Multi-text to long video generation via temporal co-denoising. arXiv preprint arXiv:2305.18264

    2023 arXiv

  61. [69]

    Wang, F.-Y.; Huang, Z.; Ma, Q.; Song, G.; Lu, X.; Bian, W.; Li, Y.; Liu, Y.; and Li, H. 2024. ZoLA: Zero-Shot Creative Long Animation Generation with Short Video Model. In European Conference on Computer Vision, 329--345. Springer

    2024

  62. [70]

    Wang, H.; Ma, C.-Y.; Liu, Y.-C.; Hou, J.; Xu, T.; Wang, J.; Juefei-Xu, F.; Luo, Y.; Zhang, P.; Hou, T.; et al. 2025 b . Lingen: Towards high-resolution minute-length text-to-video generation with linear computational complexity. In Proceedings of the Computer Vision and Patter...

    2025

  63. [71]

    Wang, J.; Sheng, H.; Cai, S.; Zhang, W.; Yan, C.; Feng, Y.; Deng, B.; and Ye, J. 2025 c . EchoShot: Multi-Shot Portrait Video Generation. arXiv preprint arXiv:2506.15838

    2025 arXiv

  64. [72]

    Wang, J.; Yuan, H.; Chen, D.; Zhang, Y.; Wang, X.; and Zhang, S. 2023 b . Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571

    2023 arXiv

  65. [73]

    Wang, Y.; Chen, X.; Ma, X.; Zhou, S.; Huang, Z.; Wang, Y.; Yang, C.; He, Y.; Yu, J.; Yang, P.; et al. 2025 d . Lavie: High-quality video generation with cascaded latent diffusion models. International Journal of Computer Vision, 133(5): 3059--3078

    2025

  66. [74]

    Wang, Y.; He, Y.; Li, Y.; Li, K.; Yu, J.; Ma, X.; Li, X.; Chen, G.; Chen, X.; Wang, Y.; et al. 2023 c . Internvid: A large-scale video-text dataset for multimodal understanding and generation. arXiv preprint arXiv:2307.06942

    2023 arXiv

  67. [75]

    Wu, Z.; Siarohin, A.; Menapace, W.; Skorokhodov, I.; Fang, Y.; Chordia, V.; Gilitschenski, I.; and Tulyakov, S. 2024. Mind the Time: Temporally-Controlled Multi-Event Video Generation. arXiv preprint arXiv:2412.05263

    2024 arXiv

  68. [76]

    Xiang, J.; Liu, G.; Gu, Y.; Gao, Q.; Ning, Y.; Zha, Y.; Feng, Z.; Tao, T.; Hao, S.; Shi, Y.; et al. 2024. Pandora: Towards general world model with natural language actions and video states. arXiv preprint arXiv:2406.09455

    2024 arXiv

  69. [77]

    Xiao, J.; Cheng, F.; Qi, L.; Gui, L.; Cen, J.; Ma, Z.; Yuille, A.; and Jiang, L. 2025 a . VideoAuteur: Towards Long Narrative Video Generation. arXiv preprint arXiv:2501.06173

    2025 arXiv

  70. [78]

    Xiao, Z.; Lan, Y.; Zhou, Y.; Ouyang, W.; Yang, S.; Zeng, Y.; and Pan, X. 2025 b . WORLDMEM: Long-term consistent world simulation with memory. arXiv preprint arXiv:2504.12369

    2025

  71. [79]

    Xie, D.; Xu, Z.; Hong, Y.; Tan, H.; Liu, D.; Liu, F.; Kaufman, A.; and Zhou, Y. 2024. Progressive autoregressive video diffusion models. arXiv preprint arXiv:2410.08151

    2024 arXiv

  72. [80]

    Xue, H.; Hang, T.; Zeng, Y.; Sun, Y.; Liu, B.; Yang, H.; Fu, J.; and Guo, B. 2022. Advancing high-resolution video-language representation with large-scale video transcriptions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 5036--5045

    2022

  73. [81]

    Yan, X.; Cai, Y.; Wang, Q.; Zhou, Y.; Huang, W.; and Yang, H. 2024. Long Video Diffusion Generation with Segmented Cross-Attention and Content-Rich Video Data Curation. arXiv preprint arXiv:2412.01316

    2024 arXiv

  74. [82]

    Yang, Z.; Teng, J.; Zheng, W.; Ding, M.; Huang, S.; Xu, J.; Yang, Y.; Hong, W.; Zhang, X.; Feng, G.; et al. 2024. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072

    2024 arXiv

  75. [83]

    Yin, S.; Wu, C.; Yang, H.; Wang, J.; Wang, X.; Ni, M.; Yang, Z.; Li, L.; Liu, S.; Yang, F.; et al. 2023. NUWA-XL: Diffusion over Diffusion for eXtremely Long Video Generation. In The 61st Annual Meeting Of The Association For Computational Linguistics

    2023

  76. [84]

    T.; Durand, F.; Shechtman, E.; and Huang, X

    Yin, T.; Zhang, Q.; Zhang, R.; Freeman, W. T.; Durand, F.; Shechtman, E.; and Huang, X. 2024. From slow bidirectional to fast autoregressive video diffusion models. arXiv preprint arXiv:2412.07772, 2

    2024

  77. [85]

    Yu, S.; Hahn, M.; Kondratyuk, D.; Shin, J.; Gupta, A.; Lezama, J.; Essa, I.; Ross, D.; and Huang, J. 2025. MALT Diffusion: Memory-Augmented Latent Transformers for Any-Length Video Generation. arXiv preprint arXiv:2502.12632

    2025 arXiv

  78. [86]

    Zhang, L.; and Agrawala, M. 2025. Packing input frame context in next-frame prediction models for video generation. arXiv preprint arXiv:2504.12626

    2025

  79. [87]

    T.; and Tan, H

    Zhang, T.; Bi, S.; Hong, Y.; Zhang, K.; Luan, F.; Yang, S.; Sunkavalli, K.; Freeman, W. T.; and Tan, H. 2025 a . Test-time training done right. arXiv preprint arXiv:2505.23884

    2025 arXiv

  80. [88]

    Zhang, Y.; Jiang, J.; Ma, G.; Lu, Z.; Huang, H.; Yuan, J.; and Duan, N. 2025 b . Generative pre-trained autoregressive diffusion transformer. arXiv preprint arXiv:2505.07344

    2025

  81. [89]

    Zhang, Y.; Xing, J.; Xia, B.; Liu, S.; Peng, B.; Tao, X.; Wan, P.; Lo, E.; and Jia, J. 2025 c . Training-Free Efficient Video Generation via Dynamic Token Carving. arXiv preprint arXiv:2505.16864

    2025

  82. [90]

    Zhao, M.; He, G.; Chen, Y.; Zhu, H.; Li, C.; and Zhu, J. 2025. Riflex: A free lunch for length extrapolation in video diffusion transformers. arXiv preprint arXiv:2502.15894

    2025 arXiv

  83. [91]

    Zheng, G.; Yuan, J.; Wang, B.; Huang, H.; Ma, G.; and Duan, N. 2025. Frame-Level Captions for Long Video Generation with Complex Multi Scenes. arXiv preprint arXiv:2505.20827

    2025 arXiv

  84. [92]

    Zheng, M.; Xu, Y.; Huang, H.; Ma, X.; Liu, Y.; Shu, W.; Pang, Y.; Tang, F.; Chen, Q.; Yang, H.; et al. 2024 a . VideoGen-of-Thought: A Collaborative Framework for Multi-Shot Video Generation. arXiv preprint arXiv:2412.02259

    2024

  85. [93]

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

    2024 arXiv

  86. [94]

    Zhou, Y.; Wang, Q.; Cai, Y.; and Yang, H. 2024 a . Allegro: Open the black box of commercial-level video generation model. arXiv preprint arXiv:2410.15458

    2024 arXiv

  87. [95]

    Zhou, Y.; Zhou, D.; Cheng, M.-M.; Feng, J.; and Hou, Q. 2024 b . Storydiffusion: Consistent self-attention for long-range image and video generation. Advances in Neural Information Processing Systems, 37: 110315--110340

    2024

  88. [96]

    Zhu, J.; Yang, H.; He, H.; Wang, W.; Tuo, Z.; Cheng, W.-H.; Gao, L.; Song, J.; and Fu, J. 2023. Moviefactory: Automatic movie creation from text using large generative models for language and images. In Proceedings of the 31st ACM International Conference on Multimedia, 9313--9319

    2023

  89. [97]

    Zhuang, S.; Huang, Z.; Zhang, Y.; Wang, F.; Fu, C.; Yang, B.; Sun, C.; Li, C.; and Wang, Y. 2025. Video-GPT via Next Clip Diffusion. arXiv preprint arXiv:2505.12489

    2025 arXiv

  90. [98]

    Zhuang, S.; Li, K.; Chen, X.; Wang, Y.; Liu, Z.; Qiao, Y.; and Wang, Y. 2024. Vlogger: Make your dream a vlog. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 8806--8817

    2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.