Pith. sign in

REVIEW 4 major objections 6 minor 7 cited by

UltraVideo: High-Quality UHD Video Dataset with Comprehensive Captions

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read UltraVideo introduces a 42K-clip native 4K/8K text-to-video dataset with 824-word structured captions and shows LoRA fine-tuning on it yields native 1K/4K generation.

desk verdict First native 4K/8K T2V dataset with structured captions, but the effectiveness claim lacks a matched control. read the letter →

arxiv 2506.13691 v1 pith:67JZNSUP submitted 2025-06-16 cs.CV

Zhucun Xue , Jiangning Zhang , Teng Hu , Haoyang He , Yinan Chen , Yuxuan Cai , Yabiao Wang , Chengjie Wang
show 3 more authors
Yong Liu Xiangtai Li Dacheng Tao
This is my paper · ORCID
classification cs.CV
keywords UltraVideoUHDdataset4K/8Ktext-to-videogenerationstructuredcaptionsLoRAfine-tuningdatacuration
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

UltraVideo proposes the first public text-to-video dataset that puts native UHD resolution first: 42K short clips (3 to 10 seconds) in 4K resolution, 22.4% of them 8K, plus 17K longer clips for future long-video work. Each clip is paired with nine structured captions and one summarized caption, averaging 824 words per video, generated almost entirely with open-source models. The paper's companion claim is that fine-tuning the small Wan-T2V-1.3B model on this data with LoRA produces UltraWan-1K/-4K, which natively generate 1K/4K videos with better text controllability than the base model, which degrades sharply when extrapolated to those resolutions. If this holds, a 42K-clip dataset can be enough to push open text-to-video models to UHD output, and the four-stage curation pipeline provides a reusable recipe for building that kind of data.

What carries the argument

The load-bearing machinery is a four-stage curation pipeline followed by a caption-sampling training strategy. Collection restricts sources to 4K/8K YouTube footage, with shots split by PySceneDetect and DINOv2 transition checks; statistical filtering removes frames with excessive text, black borders, exposure problems, or low saturation; model-based purification applies a Video Training Suitability Score threshold, RAFT motion scores, VideoCLIP-XL-v2 caption-video similarity, and Qwen2.5-VL-72B binary judgments against sixteen low-quality attributes; and captioning generates nine structured descriptions (brief, detailed, background, theme, style, shot type, camera movement, lighting, atmosphere) which Qwen3-4B folds into a summarized caption. On the model side, the key mechanism is LoRA fine-tuning of Wan-T2V-1.3B at ranks 64/16 for 1K/4K, paired with a random caption sampler that, two thirds of the time, feeds a Brief or Detailed description plus one randomly selected supplementary caption category, and one third of the time the summarized caption. This sampler lets the model see varied prompt granularities while always exposing the structured metadata.

What would settle it

Run UltraWan's identical LoRA configuration on a 42K-clip 720p or 1080p dataset with equally long structured captions and evaluate both models at 1K and 4K on the same VBench subset; if the lower-resolution dataset matches UltraWan's quality and text-consistency scores, the paper's claim that native UHD data is doing the work is not supported.

Watch

Extended reading notes

Core claim

The paper's central claim is that data, not model scale, is the bottleneck for native UHD text-to-video generation. It curates UltraVideo, a 42K-clip dataset in which every video is natively 4K or 8K and carries ten caption layers (nine structured categories plus a summary), averaging 824 words per clip. It then fine-tunes Wan-T2V-1.3B with LoRA to obtain UltraWan-1K and UltraWan-4K, and reports that these models generate 1K/4K output with higher aesthetic quality and more consistent text alignment than the same base model forced to extrapolate to 1K/4K without tuning. A human study on one-tenth of the VBench prompt set gives UltraWan-1K an 81.1% preference on video quality aesthetics over the official 480p Wan model, and the authors report that a random sample of UltraVideo clips has a 2.3% 'bad video' failure rate versus 41.5% for Koala-36M. The authors frame UltraVideo as the first public dataset that gives priority to native UHD resolution and structured captions, and they treat the UltraWan results as evidence that the dataset is doing the work.

Load-bearing premise

The paper assumes the observed 1K/4K improvement is caused by UltraVideo rather than by the fine-tuning setup, but it never trains the same LoRA recipe at the same output resolution on an equally large 720p or 1080p dataset, and the appendix's own caveats about one-epoch training and frame-count limits leave that causal role unisolated.

Editorial extensions

If this is right

  • A 42K-clip native UHD dataset is sufficient to move a 1.3B text-to-video model from failed high-resolution extrapolation to native 1K/4K generation through LoRA alone.
  • The four-stage curation pipeline, built on open-source tools, gives other groups a drop-in recipe for creating high-resolution video datasets without closed captioning APIs.
  • Random sampling among the ten caption layers means the same training video can teach both brief prompt-following and fine-grained semantic control, so annotation effort is not wasted.
  • If the reported human preferences generalize, small-scale LoRA fine-tuning on curated UHD data can rival full model fine-tuning at high resolution, making UHD generation accessible with moderate compute.
  • The released long-video subset (17K clips of 10 seconds or more) is positioned to support long-duration UHD generation research, even though it is not evaluated in this paper.

Reading between the lines

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

  • A same-size 720p or 1080p dataset with equally long structured captions, trained with the identical LoRA recipe at 1K/4K output, would settle whether native UHD pixels or the captioning/filtering pipeline drive the gains; that control is absent from the paper.
  • Because captioning relies on a specific open-source MLLM (Qwen2.5-VL-72B), the 824-word average is a snapshot of that model's current capabilities; rerunning the pipeline with a stronger captioner would likely push the average up and could shift downstream generation quality.
  • The ten caption categories (shot type, camera movement, lighting, atmosphere, and so on) read as natural conditioning axes; one testable extension is training the same LoRA setup with only one category varied to see if each axis independently controls the corresponding visual attribute.
  • The preserved native resolution and frame-rate metadata could also feed UHD super-resolution, frame interpolation, or compression research, as the paper's limitations section notes, though none of those tasks is evaluated here.
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

4 major / 6 minor

Summary. The paper introduces UltraVideo, a UHD 4K/8K text-to-video dataset containing 42K short clips (3–10s) and 17K long clips (≥10s), with nine structured caption categories plus a summarized caption averaging 824 words per video. The dataset is curated through a four-stage pipeline: source collection from 4K/8K YouTube videos, statistical filtering, model-based purification (using Qwen2.5-VL-72B, RAFT, VideoCLIP-XL-v2, VTSS), and comprehensive caption generation. To validate the dataset, the authors fine-tune Wan-T2V-1.3B with LoRA to produce UltraWan-1K and UltraWan-4K, claiming native high-resolution generation and improved text controllability, supported by VBench evaluations and a human preference study.

Significance. If the claims hold, UltraVideo is a valuable new resource: it is the first public dataset focused on native UHD (4K/8K) resolution with rich structured captions, and its curation pipeline is described in enough detail to be reproducible. The release of the dataset and models at the project page, the quantitative quality comparison against Koala-36M (2.3% vs. 41.5% human-judged failure rate), and the finding that a modest 42K-clip set can support native 1K/4K LoRA fine-tuning are all notable contributions. However, the paper's core effectiveness claim—that observed generation improvements are caused by UltraVideo's content—is not yet supported by the experimental design.

major comments (4)
  1. [Sec. 4, Table 4; Sec. 3.2] The central claim that UltraWan's improvements are due to the UltraVideo dataset is not isolated: there is no same-resolution control trained with the same LoRA recipe and hyperparameters on an alternative dataset (e.g., Koala-36M, MiraData, OpenVidHD, or a lower-quality subset of UltraVideo itself). Without such an ablation, the gains could stem from the fine-tuning setup (resolution, caption sampling strategy, training budget) rather than from the dataset content. This is load-bearing for the stated contribution, 'demonstrating the effectiveness of our data curation.'
  2. [Sec. 4, Table 4; Table 5] The VBench results do not support the claim of 'more consistent text controllability': UltraWan-1K (LoRA) has an Overall Consistency of 23.99%, lower than the official Wan-T2V-1.3B-480p score of 25.31%. The human study in Table 5 shows only 54.5% preference for text consistency from 10 subjects, which is not statistically analyzed. Additionally, the human study compares UltraWan-1K at 1K against official Wan at 480p, entangling resolution with training data. A larger human study and reporting of full VBench results with confidence intervals are needed before the controllability claim can be accepted.
  3. [Sec. 4] The evaluation uses only one-tenth of VBench prompts (~96 prompts) and the 4K model's results are reported after downsampling to 1K on some dimensions (Motion Smoothness, Dynamic Degree). With ~96 prompts, per-dimension scores are noisy (e.g., Spatial Relationship 0.00–36.66%, Multiple Objects 25.00–82.29%). The paper should state how the subset was sampled and why it is representative, and either run the full benchmark or explicitly frame the results as a pilot evaluation.
  4. [Appendix D; Table 3] The paper concedes that UltraWan-4K may be undertrained (one epoch) and exhibits more artifacts, and Table 3 shows the 4K model is trained with only 29 frames versus 81 for 1K. This training-budget discrepancy further confounds the dataset's causal role, since the 4K result does not reflect a controlled comparison at matched training budget. Either provide a better-trained 4K model or temper the claim that the dataset enables native 4K generation; as written, the 4K demonstration is not strong enough to validate the dataset's effectiveness at UHD resolution.
minor comments (6)
  1. [Sec. 2.1, Reference [5]] Reference [5] for PySceneDetect contains the placeholder '[Insert Date]' in the URL; please provide the actual access date or a stable DOI.
  2. [Sec. 2.2] The filtering thresholds (2% text-area ratio, 5% problematic-frame ratio, brightness thresholds of 250/5, 12% pixel ratio, variance 1.2, etc.) are all presented as fixed values without sensitivity analyses; a short discussion of how these were chosen and their robustness would strengthen the curation pipeline description.
  3. [Sec. 2.5] The human failure-rate comparison with Koala-36M should state whether the same evaluation checklist and sampling procedure were used for both datasets, and whether the 1000-video samples were randomized and blinded to the evaluators; this would make the 2.3% vs. 41.5% result more credible.
  4. [Sec. 3.2] The subsection title 'Memory-efficient HDR plugins of Wan-1K/-4K LoRA' appears to be a misnomer—the text describes LoRA for resolution scaling, not HDR (high dynamic range); please correct the wording.
  5. [Abstract; Sec. 2.4] The abstract and introduction state that each video has '10 types of semantic tags' and the contribution list mentions '10 types', but Sec. 2.4 enumerates 9 structured categories plus a summarized caption; clarify whether the summarized caption counts as the 10th type.
  6. [Table 3; Figure 1] In Table 3, the 'GPU Hours' entries '3.4K' and '7.6K' are ambiguous (3,400 and 7,600 hours, presumably); specify the units explicitly. Also, the Figure 1 caption contains a stray period and the phrase 'scenes . with the right side'—please fix the typo.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the dataset and model claims are empirical and evaluated against external benchmarks, with no load-bearing self-citation or fitted-input-as-prediction.

full rationale

This paper's central claims are empirical artifacts: a curated UHD video dataset and a LoRA fine-tuned video generation model. The curation pipeline uses objective resolution, frame-rate, and statistical filters as well as external models (Qwen2.5-VL, VideoCLIP-XL-v2, RAFT), and the dataset itself is released for independent inspection. The effectiveness claim for UltraWan is evaluated on external VBench prompts and human preference judgments rather than on the training captions alone, so the reported results are not forced by construction from the training data. The paper does not invoke a uniqueness theorem, nor does it rely on load-bearing self-citations; prior datasets and models such as Koala-36M, MiraData, Wan, and Qwen are external works. The main weaknesses are experimental-design confounds: there is no same-resolution control trained on an alternative dataset, and the human study compares UltraWan-1K at 1K with official Wan at 480p, entangling resolution and training data. These are correctness and validity concerns, not circularity. Using the same MLLM for caption generation and attribute judgment could introduce bias, but it does not make the paper's claims equivalent to its inputs by definition. No circular step can be exhibited with a specific equation or constructed equivalence.

Assumptions & free parameters 9 free parameters · 5 assumptions · 0 invented entities

The dataset curation and model fine-tuning involve several hand-chosen thresholds and assumptions about the reliability of automated models. These are not fitted to a target result, but they shape the dataset and therefore the conclusions.

free parameters (9)
  • text_area_ratio_threshold = 2%
    Frames with more than 2% text area are considered problematic; hand-chosen in Sec. 2.2.
  • problematic_frame_ratio_threshold = 5%
    Videos with over 5% problematic frames are excluded; hand-chosen in Sec. 2.2.
  • black_border_mean_threshold = 3
    Frames with border mean value below 3 are flagged; hand-chosen in Sec. 2.2.
  • exposure_pixel_ratio = 12%
    Frames where >12% of pixels are above 250 or below 5 are flagged; hand-chosen in Sec. 2.2.
  • graying_variance_threshold = 1.2
    Frames with average RGB variance below 1.2 are flagged; hand-chosen in Sec. 2.2.
  • vtss_threshold = 0.01
    Videos with VTSS score below 0.01 are removed; threshold chosen in Sec. 2.3 (scale from Koala-36M).
  • motion_score_interval = [0.1, 100]
    Retain videos with global average RAFT motion in this interval; hand-chosen in Sec. 2.3.
  • caption_similarity_threshold = 0.2
    Filter video-caption pairs below this VideoCLIP-XL-v2 similarity; hand-chosen in Sec. 2.3.
  • caption_sampling_probability = 1/3
    Random caption sampling strategy picks among three description types with probability 1/3; design choice in Sec. 3.2.
assumptions (5)
  • domain assumption YouTube 4K/8K videos, selected via keyword search and meta-data filtering, are representative of diverse, high-quality UHD content.
    Invoked in Sec. 2.1 to justify the source pool; no external validation of representativeness.
  • domain assumption Qwen2.5-VL-72B provides accurate judgments for low-quality attribute detection and caption generation.
    Used in Sec. 2.3 and 2.4; no human agreement analysis for the MLLM's binary judgments.
  • domain assumption VideoCLIP-XL-v2 similarity scores are a valid measure of video-caption consistency for filtering.
    Used in Sec. 2.3; the metric is adopted without calibration against human judgments.
  • domain assumption VBench metrics and a 10-subject human preference study are valid proxies for generation quality at 1K/4K.
    Sec. 4 relies on these for the central claim; the paper itself notes VBench is inaccurate for high-resolution evaluation.
  • ad hoc to paper The hand-chosen filtering thresholds (e.g., 2% text, 5% problematic frames) generalize across diverse video content.
    Thresholds in Sec. 2.2 are set without sensitivity analysis or justification from data distributions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of UltraVideo: High-Quality UHD Video Dataset with Comprehensive Captions." pith.science (2026). https://pith.science/paper/67JZNSUP

@misc{pith2026250613691,
  author       = {Pith},
  title        = {Pith review of: UltraVideo: High-Quality UHD Video Dataset with Comprehensive Captions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/67JZNSUP}},
  note         = {Machine review of arXiv:2506.13691}
}
read the original abstract

The quality of the video dataset (image quality, resolution, and fine-grained caption) greatly influences the performance of the video generation model. The growing demand for video applications sets higher requirements for high-quality video generation models. For example, the generation of movie-level Ultra-High Definition (UHD) videos and the creation of 4K short video content. However, the existing public datasets cannot support related research and applications. In this paper, we first propose a high-quality open-sourced UHD-4K (22.4\% of which are 8K) text-to-video dataset named UltraVideo, which contains a wide range of topics (more than 100 kinds), and each video has 9 structured captions with one summarized caption (average of 824 words). Specifically, we carefully design a highly automated curation process with four stages to obtain the final high-quality dataset: \textit{i)} collection of diverse and high-quality video clips. \textit{ii)} statistical data filtering. \textit{iii)} model-based data purification. \textit{iv)} generation of comprehensive, structured captions. In addition, we expand Wan to UltraWan-1K/-4K, which can natively generate high-quality 1K/4K videos with more consistent text controllability, demonstrating the effectiveness of our data curation.We believe that this work can make a significant contribution to future research on UHD video generation. UltraVideo dataset and UltraWan models are available at https://xzc-zju.github.io/projects/UltraVideo.

Figures

Figures reproduced from arXiv: 2506.13691 by the authors.

Figure 1
Figure 1. UltraVideo has higher visual quality and ultra-high resolution (≥ 4K), along with comprehensive structured captions (bottom). Compared with current text-to-video (T2V) datasets, it can meet the growing demand for native high-resolution T2V applications. Benefiting from the carefully designed curation process, this dataset contains diverse visually attractive scenes. with the right side showing random samples with di… view at source ↗
Figure 2
Figure 2. Wan-T2V-1.3B [34] shows a significant decline in visual quality and semantic consistency as the resolution increases, and it fails to generate high-resolution videos without. 2 Curating UltraVideo Dataset Recent T2V datasets emphasize the quantity of videos (million-level 720p videos) with detailed captions that can support the pre-training of video models. In contrast, we mainly focus on the quality of the UHD vide… view at source ↗
Figure 3
Figure 3. Our video curation process that includes four data collection processes: ➀ Video Clips Collection (Sec. 2.1), ➁ Statistical Data Filtering (Sec. 2.2), ➂ Model-based data purification (Sec. 2.3), and ➃ Comprehensive Structured Caption (Sec. 2.4). Ultimately, we obtain 42K high￾quality UHD short clips with comprehensive descriptions. Frame number filtering. Mainstream video generation models only support video generat… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Statistical distributions of our UltraVideo from different perspectives. 3.1 Resolution Scaling of Wan. Powerless extrapolation. Benefiting from the relative position encoding and rotational invariance of RoPE, the DiT-based Wan has a certain degree of variable resolut…
Figure 5
Figure 5. Figure 5: Intuitive results with the prompt in VBench [ [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Our UltraWan-1K is capable of generating semantically consistent videos. [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 7 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. PhiZero: A World Model Built Around Physical Language

    cs.CV 2026-07 conditional novelty 6.0 of 10

    A self-supervised discrete physical-language bottleneck plus a VLM reasoner lets a world model predict state transitions before rendering video, improving physical coherence and enabling zero-shot motion transfer.

  2. FlashDecoder: Real-Time Latent-to-Pixel Streaming Decoder with Transformers

    cs.CV 2026-07 conditional novelty 6.0 of 10

    A pure-Transformer streaming decoder with a fixed-size rolling KV cache decodes video latents 3.6–4.7x faster than 3D convolutional decoders with comparable PSNR.

  3. LUVE : Latent-Cascaded Ultra-High-Resolution Video Generation with Dual Frequency Experts

    cs.CV 2026-02 conditional novelty 6.0 of 10

    A latent-cascaded video generation framework with dual frequency-split experts reports state-of-the-art 2K/4K video generation on VBench, FIDpatch, and human preference.

  4. InverseCrafter: Efficient Video ReCapture as a Latent Domain Inverse Problem

    cs.CV 2025-12 conditional novelty 6.0 of 10

    A training-free, near-zero-overhead inverse solver for novel-view video generation and inpainting that projects masks into continuous multi-channel latent masks and applies DDS with conjugate gradient in latent space.

  5. Phase-Aligned RoPE for Mixed-Resolution Diffusion Transformer

    cs.CV 2025-11 conditional novelty 6.0 of 10

    Expressing all RoPE positions on the query's grid ('one attention, one scale') plus a small boundary content-exchange step restores mixed-resolution diffusion generation that naive position interpolation destroys.

  6. Your Data Manifold is Secretly a Reward Model: Shell-LCC for Text-to-Video Generation

    cs.CV 2026-06 unverdicted novelty 5.0 of 10

    Shell-LCC models the high-quality data manifold as an isotropic shell to derive cost-free reward signals that improve realism and high-frequency details in text-to-video generation.

  7. Effectively obtaining acoustic, visual and textual data from videos

    cs.MM 2025-09 conditional novelty 4.0 of 10

    A video-processing pipeline created a 2.24 million-sample audio-image-text dataset, with text captions generated by BLIP from video frames.

Reference graph

Works this paper leans on

45 extracted references · 31 canonical work pages · cited by 7 Pith papers

  1. [1]

    L. AI. aesthetic-predictor. https://github.com/LAION-AI/aesthetic-predictor , 2022. 14

    work page 2022

  2. [2]

    S. Bai, K. Chen, X. Liu, J. Wang, W. Ge, S. Song, K. Dang, P. Wang, S. Wang, J. Tang, et al. Qwen2.5-vl technical report. arXiv preprint arXiv:2502.13923, 2025. 2, 5

    arXiv 2025

  3. [3]

    M. Bain, A. Nagrani, G. Varol, and A. Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In ICCV, 2021. 6, 13

    work page 2021

  4. [4]

    Blattmann, T

    A. Blattmann, T. Dockhorn, S. Kulal, D. Mendelevitch, M. Kilian, D. Lorenz, Y . Levi, Z. English, V . V oleti, A. Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 13, 14

    arXiv 2023

  5. [5]

    Castellano

    B. Castellano. Pyscenedetect: Python-based video scene detector, March 2024. URL https: //github.com/Breakthrough/PySceneDetect. Accessed: [Insert Date]. 3

    work page 2024

  6. [6]

    Chen and W

    D. Chen and W. Dolan. Collecting highly parallel data for paraphrase evaluation. In ACL, 2011. 13

    work page 2011

  7. [7]

    H. Chen, Y . Zhang, X. Cun, M. Xia, X. Wang, C. Weng, and Y . Shan. Videocrafter2: Overcoming data limitations for high-quality video diffusion models. In CVPR, 2024. 13

    work page 2024

  8. [8]

    T.-S. Chen, A. Siarohin, W. Menapace, E. Deyneka, H.-w. Chao, B. E. Jeon, Y . Fang, H.-Y . Lee, J. Ren, M.-H. Yang, et al. Panda-70m: Captioning 70m videos with multiple cross-modality teachers. In CVPR, 2024. 2, 3, 6, 13

    work page 2024

Show all 45 references
  1. [9]

    H. Deng, T. Pan, H. Diao, Z. Luo, Y . Cui, H. Lu, S. Shan, Y . Qi, and X. Wang. Autoregressive video generation without vector quantization. arXiv preprint arXiv:2412.14169, 2024. 13

    2024 arXiv

  2. [10]

    I. J. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y . Bengio. Generative adversarial nets. InNeurIPS, 2014. 13

    2014

  3. [11]

    Y . Guo, C. Yang, A. Rao, Z. Liang, Y . Wang, Y . Qiao, M. Agrawala, D. Lin, and B. Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. In ICLR, 2024. 13

    2024

  4. [12]

    J. Ho, A. Jain, and P. Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 13

    2020

  5. [13]

    Huang, Y

    Z. Huang, Y . He, J. Yu, F. Zhang, C. Si, Y . Jiang, Y . Zhang, T. Wu, Q. Jin, N. Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In CVPR, 2024. 8, 16

    2024

  6. [14]

    X. Ju, Y . Gao, Z. Zhang, Z. Yuan, X. Wang, A. Zeng, Y . Xiong, Q. Xu, and Y . Shan. Miradata: A large-scale video dataset with long durations and structured captions. NeurIPS, 2024. 3, 4, 5, 6, 13

    2024

  7. [15]

    W. Kong, Q. Tian, Z. Zhang, R. Min, Z. Dai, J. Zhou, J. Xiong, X. Li, B. Wu, J. Zhang, et al. Hunyuanvideo: A systematic framework for large video generative models. arXiv preprint arXiv:2412.03603, 2024. 3, 4, 13

    2024 arXiv

  8. [16]

    B. Lin, Y . Ge, X. Cheng, Z. Li, B. Zhu, S. Wang, X. He, Y . Ye, S. Yuan, L. Chen, et al. Open-sora plan: Open-source large video generation model. arXiv preprint arXiv:2412.00131, 2024. 6, 13

    2024 arXiv

  9. [17]

    Loshchilov and F

    I. Loshchilov and F. Hutter. Decoupled weight decay regularization. In ICLR, 2019. 8

    2019

  10. [18]

    G. Ma, H. Huang, K. Yan, L. Chen, N. Duan, S. Yin, C. Wan, R. Ming, X. Song, X. Chen, et al. Step-video-t2v technical report: The practice, challenges, and future of video foundation model. arXiv preprint arXiv:2502.10248, 2025. 13

    2025 arXiv

  11. [19]

    Miech, D

    A. Miech, D. Zhukov, J.-B. Alayrac, M. Tapaswi, I. Laptev, and J. Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In ICCV, 2019. 6, 13

    2019

  12. [20]

    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. In ICLR, 2025. 2, 3, 6, 13 10

    2025

  13. [21]

    C. Ng, D. Schnurr, E. Luhman, J. Taylor, L. Jing, N. Summers, R. Wang, R. Sahai, R. O’Rourke, T. Luhman, W. DePue, and Y . Guo. Sora | openai, 2024. 13

    2024

  14. [22]

    Paddleocr

    PaddlePaddle. Paddleocr. https://github.com/PaddlePaddle/PaddleOCR, 2023. 4

    2023

  15. [23]

    Podell, Z

    D. Podell, Z. English, K. Lacey, A. Blattmann, T. Dockhorn, J. Müller, J. Penna, and R. Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. In ICLR, 2024. 13

    2024

  16. [24]

    Radford, J

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 14

    2021

  17. [25]

    Rombach, A

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 13

    2022

  18. [26]

    Magi-1: Autoregressive video generation at scale, 2025

    Sand-AI. Magi-1: Autoregressive video generation at scale, 2025. URL https://static. magi.world/static/files/MAGI_1.pdf. 13

    2025

  19. [27]

    Schuhmann, R

    C. Schuhmann, R. Vencu, R. Beaumont, R. Kaczmarczyk, C. Mullis, A. Katta, T. Coombes, J. Jitsev, and A. Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 13

    2021 arXiv

  20. [28]

    Schuhmann, R

    C. Schuhmann, R. Beaumont, R. Vencu, C. Gordon, R. Wightman, M. Cherti, T. Coombes, A. Katta, C. Mullis, M. Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in neural information processing systems , 35: 25278–2529...

    2022

  21. [29]

    Soomro, A

    K. Soomro, A. R. Zamir, and M. Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 13

    2012 arXiv

  22. [30]

    Streamlit

    Streamlit. Streamlit. https://github.com/streamlit/streamlit, 2018-2025. 9

    2018

  23. [31]

    Z. Tan, X. Yang, L. Qin, and H. Li. Vidgen-1m: A large-scale dataset for text-to-video generation. arXiv preprint arXiv:2408.02629, 2024. 3, 6, 13, 14

    2024 arXiv

  24. [32]

    Q. Team. Qwen3, April 2025. URL https://qwenlm.github.io/blog/qwen3/. 5

    2025

  25. [33]

    Teed and J

    Z. Teed and J. Deng. Raft: Recurrent all-pairs field transforms for optical flow. In ECCV, 2020. 5

    2020

  26. [34]

    A. Wang, B. Ai, B. Wen, C. Mao, C.-W. Xie, D. Chen, F. Yu, H. Zhao, J. Yang, J. Zeng, J. Wang, J. Zhang, J. Zhou, J. Wang, J. Chen, K. Zhu, K. Zhao, K. Yan, L. Huang, M. Feng, N. Zhang, P. Li, P. Wu, R. Chu, R. Feng, S. Zhang, S. Sun, T. Fang, T. Wang, T. Gui, T. Weng, T. Shen...

    2025 arXiv

  27. [35]

    J. Wang, C. Wang, K. Huang, J. Huang, and L. Jin. VideoCLIP-XL: Advancing long description understanding for video CLIP models. In EMNLP, 2024. 5

    2024

  28. [36]

    Q. Wang, Y . Shi, J. Ou, R. Chen, K. Lin, J. Wang, B. Jiang, H. Yang, M. Zheng, X. Tao, et al. Koala-36m: A large-scale video dataset improving consistency between fine-grained conditions and video content. arXiv preprint arXiv:2410.08260, 2024. 2, 3, 5, 6, 14

    arXiv

  29. [37]

    Wang and Y

    W. Wang and Y . Yang. Videoufo: A million-scale user-focused dataset for text-to-video generation. arXiv preprint arXiv:2503.01739, 2025. 6

    2025 arXiv

  30. [38]

    Y . Wang, Y . He, Y . Li, K. Li, J. Yu, X. Ma, X. Li, G. Chen, X. Chen, Y . Wang, P. Luo, Z. Liu, Y . Wang, L. Wang, and Y . Qiao. Internvid: A large-scale video-text dataset for multimodal understanding and generation. In ICLR, 2024. 6, 13 11

    2024

  31. [39]

    Y . Wang, T. Xiong, D. Zhou, Z. Lin, Y . Zhao, B. Kang, J. Feng, and X. Liu. Loong: Generating minute-level long videos with autoregressive language models.arXiv preprint arXiv:2410.02757,

    arXiv

  32. [40]

    J. Z. Wu, Y . Ge, X. Wang, S. W. Lei, Y . Gu, Y . Shi, W. Hsu, Y . Shan, X. Qie, and M. Z. Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In ICCV,

  33. [41]

    Xiong, Y

    T. Xiong, Y . Wang, D. Zhou, Z. Lin, J. Feng, and X. Liu. Lvd-2m: A long-take video dataset with temporally dense captions. arXiv preprint arXiv:2410.10816, 2024. 6, 13, 14

    2024 arXiv

  34. [42]

    J. Xu, T. Mei, T. Yao, and Y . Rui. Msr-vtt: A large video description dataset for bridging video and language. In CVPR, 2016. 13

    2016

  35. [43]

    H. Xue, T. Hang, Y . Zeng, Y . Sun, B. Liu, H. Yang, J. Fu, and B. Guo. Advancing high-resolution video-language representation with large-scale video transcriptions. In CVPR, 2022. 6, 13

    2022

  36. [44]

    Z. Yang, J. Teng, W. Zheng, M. Ding, S. Huang, J. Xu, Y . Yang, W. Hong, X. Zhang, G. Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 13

    2024 arXiv

  37. [45]

    Zellers, X

    R. Zellers, X. Lu, J. Hessel, Y . Yu, J. S. Park, J. Cao, A. Farhadi, and Y . Choi. Merlot: Multimodal neural script knowledge models. Advances in neural information processing systems, 34:23634–23651, 2021. 13 12 Appendix The supplementary material presents the following sect...

    2021

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.