Pith. sign in

REVIEW 4 major objections 6 minor 3 cited by

Phantom-Data : Towards a General Subject-Consistent Video Generation Dataset

T0 review · 4 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 A new dataset claims to fix the copy-paste problem in subject-to-video generation by training on roughly one million identity-consistent, cross-context pairs.

desk verdict A genuinely useful large-scale cross-pair dataset for subject-to-video generation, but the headline identity-consistency claim is not supported by the paper's own numbers. read the letter →

arxiv 2506.18851 v1 pith:W3NEJD2Q submitted 2025-06-23 cs.CV

Zhuowei Chen , Bingchuan Li , Tianxiang Ma , Lijie Liu , Mingcong Liu , Yi Zhang , Gen Li , Xinghui Li
show 3 more authors
Siyu Zhou Qian He Xinglong Wu
This is my paper
classification cs.CV
keywords Phantom-Datasubject-to-videogenerationcross-pairtrainingcopy-pasteproblemidentityconsistencytext-promptalignmentvideodatasetconstructionvisualgrounding
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

Subject-to-video generation models tend to copy the reference image's background and pose instead of following what the text prompt asks for. This paper claims the cause is the in-pair training paradigm, in which reference and target frames come from the same video, and presents Phantom-Data, the first general-purpose cross-pair subject-to-video consistency dataset, with roughly one million identity-consistent pairs in which the same subject appears in different contexts. The paper argues that training on such pairs decouples identity from context, and reports that a 1.3-billion-parameter subject-to-video model trained on Phantom-Data improves prompt alignment and video quality while keeping identity consistency at the same level as in-pair baselines. A user study in the paper also reports a 76% preference for the cross-pair-trained outputs.

What carries the argument

The load-bearing mechanism is a three-stage construction pipeline: an open-vocabulary subject detector that uses a vision-language model to ground caption nouns onto complete, specific visual regions; a contextually diverse retrieval stage over 53 million video clips and 3 billion images with category-specific identity embeddings, using both lower and upper similarity bounds to exclude duplicates; and prior-guided identity verification, which for living subjects restricts candidates to different clips of the same long video and for products requires visible logos, followed by a vision-language pairwise check for same identity plus different background. This pipeline is what turns raw footage and images into the one-million-pair training set claimed to break the context-identity entanglement.

What would settle it

A random sample of Phantom-Data pairs could be sent to human annotators to judge two questions: does the reference image show the same identity as the target video, and do the two contexts differ. If a substantial share of pairs fails either check, or if a control model trained on the same one million pairs with the verification stage removed matches the reported gains, the claim that identity-consistent cross-pair data is the active ingredient would collapse.

Watch

Extended reading notes

Core claim

The central claim is that cross-pair supervision—identity-consistent reference and target samples drawn from different scenes—is enough to mitigate the copy-paste problem in general subject-to-video generation. According to the paper, Phantom-Data is the first open general-purpose dataset built this way, covering humans, animals, products, environments, clothing, and multi-subject scenes. The experiments show that models trained on it surpass in-pair, augmented in-pair, and face-only cross-pair baselines on prompt-following and visual-quality metrics while matching in-pair identity consistency; the reported reward-based text-alignment score roughly doubles relative to the in-pair baseline. The paper interprets this as evidence that contextual diversity in training pairs, not just data volume, is what lets the model follow textual instructions without losing the subject.

Load-bearing premise

The pipeline assumes its automated identity verification is trustworthy at scale—for living subjects, that different clips from one long video show the same individual, and for products, that a visible brand logo identifies the exact item; if false positives slip through, the reported gains could come from sheer data volume and context variety rather than from genuine identity-consistent cross-pair supervision.

Editorial extensions

If this is right

  • If the central claim is correct, subject-to-video models trained on cross-pair data should honor prompts that move the subject to new backgrounds, actions, and camera settings rather than echoing the reference frame.
  • The pipeline extends cross-pair supervision beyond faces to general categories; the ablations indicate that adding humans, animals, products, and multi-subject scenes each improves consistency or prompt following.
  • Scale matters: the 100k-to-1M comparison shows that larger cross-pair datasets keep improving generation quality, so the dataset's scale is part of its claimed benefit.
  • Face-only cross-pair training is not sufficient; the face-only baseline trails the full-object cross-pair model on both subject consistency and prompt following, suggesting domain coverage is a necessary component.
  • The reported 76% user preference for the cross-pair model over all baselines implies the gains are visible to human viewers, not just measurable with automated metrics.

Reading between the lines

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

  • If the in-pair entanglement diagnosis is right, the same failure mode should appear in other reference-conditioned generation tasks, such as single-image animation or multi-concept editing; analogous cross-pair datasets could transfer the cure.
  • Because the pipeline requires visible logos for product identity, the dataset is likely thin on unbranded or generic objects; an extension using instance-level feature matching could widen product coverage.
  • Restricting living-subject verification to clips from the same long-form video bounds the contextual shift for people and animals; mining true same-identity pairs across independent recordings is the natural next step toward fully context-free identity.
  • The verification threshold is a precision-recall dial: tightening it should raise identity fidelity and lower scale, and one testable extension is to measure downstream prompt-alignment gains as a function of that dial.
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 manuscript introduces Phantom-Data, a proposed dataset of approximately one million cross-pair reference-subject/target-video pairs for subject-consistent video generation. The construction pipeline is described in three stages: vision-language based open-set subject detection, retrieval of candidates from a bank of 53M videos and 3B images, and prior-guided identity verification using category-specific constraints and a VLM pairwise check. The authors validate the dataset by training a 1.3B-parameter Phantom-wan model on in-pair, augmented in-pair, face cross-pair, and Phantom-Data settings, and report that the proposed data improves prompt following and video quality while keeping identity consistency comparable. Experiments include quantitative comparisons on 100 test prompts, ablations on subject diversity and scale, qualitative examples, and a small user study.

Significance. If the identity-consistency guarantee of the pipeline could be established, the dataset would be a valuable community resource: it is the first open general-purpose cross-pair S2V dataset, it addresses the copy-paste problem with real instead of synthetic cross-context pairs, and its scale (about 1M pairs) is suitable for training modern video diffusion models. The paper also provides useful ablations showing that subject diversity and data scale matter. However, the current evidence is insufficient to support the two halves of the central claim: the verification stage is not validated quantitatively, and the reported DINO metric partially contradicts the 'on par' identity claim. The contribution is therefore promising but not yet substantiated.

major comments (4)
  1. [§5.2, Table 2] Table 2 directly bears on the abstract's claim that identity consistency is preserved 'on par' with in-pair baselines, and the reported DINO subject-consistency score contradicts that claim: Ours is 0.416 versus 0.478 for In-pair, a relative drop of about 13%. The evaluation is based on only 100 test cases and a single training run per condition, with no confidence intervals, bootstrap estimates, or repeated seeds, so it is unclear whether the gap is noise or a real identity-preservation cost. Please report uncertainty (e.g., bootstrap CIs or multiple seeds) and either temper the 'on par' claim or provide evidence that the DINO gap is acceptable (for example, by showing that the drop is concentrated in categories where identity is less critical).
  2. [§4.2.3] The prior-guided verification stage is the only mechanism that guarantees identity consistency in the final dataset, but its accuracy is never quantified. Living-subject candidates are restricted to different clips of the same long-form video, product candidates must contain a visible brand logo, and a VLM pairwise check is applied, yet no precision, recall, false-positive rate, or human-agreement numbers are reported for any of these filters. Without such numbers, the claim that Phantom-Data contains approximately one million identity-consistent pairs is unverified; the downstream improvements could in principle come from the larger and more diverse corpus rather than from correct identity pairing. Please add a human-evaluated precision study on a random sample of final pairs, reported per subject category and per filter stage.
  3. [§5.3, Tables 3 and 4] The ablations show that data scale and subject diversity improve the metrics, but they do not isolate the contribution of the identity-verification stage or of cross-pair identity correctness. In particular, the 100k-to-1M scale comparison changes Reward-TA by only 0.031 (3.796 to 3.827) and DINO by 0.008 (0.408 to 0.416), so the claimed gains are not clearly attributable to the verification pipeline. Please compare against a control that trains on the same number of retrieval candidates without the VLM verification filter (or with deliberately corrupted pairings), so the reader can see what the verification stage contributes beyond scale and contextual diversity.
  4. [§5.1 and Appendix B] The main evaluation uses 100 hand-written prompts with one training run per method, and the user study uses six participants who each evaluate 50 groups, with no inter-annotator agreement or significance test reported. This is a thin empirical basis for the general claim that Phantom-Data 'significantly improves' prompt alignment and visual quality across all subject categories. Please add per-category breakdowns, more than one seed (or at least bootstrap confidence intervals), and report the user-study protocol in enough detail to judge its validity.
minor comments (6)
  1. [§5.1] The text says subject-video consistency is evaluated using CLIP [13], but reference [13] is the PuLID paper, not the original CLIP paper; this citation should be corrected to [34] or an appropriate evaluation protocol.
  2. [§5.3, Table 3] Table 3 is garbled in the text: the row '+ IP/animal' is undefined, and the formatting of the values makes the incremental deltas hard to read; please re-typeset the table and define the abbreviation.
  3. [§4.2.1, Eq. (3)] V_person is defined as the concatenation of E_IR(I) and E_arcface(I_face); please state how the face crop I_face is obtained and how the two embedding dimensions are aligned.
  4. [§4.2.2] The upper and lower similarity thresholds for retrieval are described qualitatively but never reported; please give the values or a sensitivity analysis, since these thresholds directly influence the diversity/identity trade-off.
  5. [§5.2] The sentence 'The Face cross-pair method performs slightly better on prompt following' is ambiguous; it should say 'than the in-pair baselines' to avoid implying it beats the proposed method (Table 2 shows Reward-TA 3.022 vs. 3.827).
  6. [Appendix B] The user study should state whether the four videos in each group were anonymized and randomized, and should report the number of videos per category; the current six-participant design is too small to support the 76% claim without additional detail.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the dataset pipeline and downstream evaluation are independent; self-citations to Phantom-wan and PuLID are not load-bearing.

full rationale

The paper's derivation chain is: construct cross-pair subject-to-video data via detection, retrieval, and verification, then train an open-source S2V model on that data and compare against in-pair baselines. I find no step where a claimed result is defined as its own input, nor any fitted parameter renamed as a prediction. The retrieval stage uses ArcFace, an external CLIP-based object embedder, and category priors, while the downstream evaluation uses DINO, GPT-4o, Reward-TA, and VBench; these are not the same functions used to select pairs, so no metric is forced by construction. The VLM-based verification is a data-quality filter with thresholds applied during dataset construction, not a parameter fitted to the 100-case evaluation suite. The paper does cite the authors' own Phantom model [27] as the training backbone and PuLID [13] as a CLIP-related reference, but these are testbeds and standard metrics rather than premises that determine the conclusion; the central claim that cross-pair training improves prompt alignment and visual quality is an empirical result with independent quantitative content. Concerns that the 'on par' identity statement conflicts with the DINO gap (0.416 vs 0.478) or that the VLM verification precision is unreported are correctness and robustness issues, not circularity. Under the stated rules, this is a normal, non-circular dataset paper.

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

No new physical entities are introduced. The central claim depends on hand-set thresholds and on several domain assumptions about detection, retrieval, and VLM verification, which are listed above.

free parameters (3)
  • Bounding box acceptance criteria = 4%-90% image area, min 128x128, IoU > 0.8 suppressed
    Hand-set thresholds in Step 4 (Bbox Filtering) of Section 4.2.1 control which detected subjects enter the retrieval bank; no sensitivity analysis is reported.
  • Retrieval similarity thresholds = not reported
    Section 4.2.2 applies upper and lower similarity bounds to discard duplicates and unrelated identities; the actual values are not given, so the identity-diversity trade-off is not reproducible.
  • Frame sampling timestamps = t=0.05, 0.5, 0.95
    Step 1 in Section 4.2.1 samples three frames at these timestamps, following [6]; this choice affects which subjects are detected and matched.
assumptions (5)
  • domain assumption In-pair training entangles identity with background and causes copy-paste.
    Motivation in Section 1 and Figure 2; the entire cross-pair design depends on this diagnosis.
  • domain assumption Living subjects in different clips of the same long-form video are the same identity.
    Section 4.2.3 'Prior-Based Identity Verification' uses this to guarantee identity for humans and animals.
  • domain assumption VLM pairwise verification reliably detects identity consistency and contextual diversity at scale.
    Section 4.2.3 'VLM-Based Consistency Verification'; no false positive/negative rates are reported.
  • domain assumption Fine-tuned CLIP embeddings (Eq. 2) preserve identity for general objects.
    Section 4.2.2 'Subject Representation'; retrieval quality for products and animals depends on this encoder.
  • domain assumption LAION images contain matching identities for detected subjects.
    Section 4.2.2 'Large-Scale Retrieval Bank' augments the bank with 3B images; identity match rate in this cross-source setting is not measured.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Phantom-Data : Towards a General Subject-Consistent Video Generation Dataset." pith.science (2026). https://pith.science/paper/W3NEJD2Q

@misc{pith2026250618851,
  author       = {Pith},
  title        = {Pith review of: Phantom-Data : Towards a General Subject-Consistent Video Generation Dataset},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/W3NEJD2Q}},
  note         = {Machine review of arXiv:2506.18851}
}
read the original abstract

Subject-to-video generation has witnessed substantial progress in recent years. However, existing models still face significant challenges in faithfully following textual instructions. This limitation, commonly known as the copy-paste problem, arises from the widely used in-pair training paradigm. This approach inherently entangles subject identity with background and contextual attributes by sampling reference images from the same scene as the target video. To address this issue, we introduce \textbf{Phantom-Data, the first general-purpose cross-pair subject-to-video consistency dataset}, containing approximately one million identity-consistent pairs across diverse categories. Our dataset is constructed via a three-stage pipeline: (1) a general and input-aligned subject detection module, (2) large-scale cross-context subject retrieval from more than 53 million videos and 3 billion images, and (3) prior-guided identity verification to ensure visual consistency under contextual variation. Comprehensive experiments show that training with Phantom-Data significantly improves prompt alignment and visual quality while preserving identity consistency on par with in-pair baselines.

Discussion (0). Sign in to comment.

Forward citations

Cited by 3 Pith papers

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

  1. FlowMimic: Mask-free Visual Editing and Generation with Pixel-pair Warped Flow Field for Online Video Editing Data Generation and Modality Mimicry

    cs.CV 2026-07 conditional novelty 6.0 of 10

    Video editing can be learned from image-edit pairs that are synthetically warped into videos, plus self-distillation losses that align image and video outputs.

  2. Rethinking Position Embedding as a Context Controller for Multi-Reference and Multi-Shot Video Generation

    cs.CV 2026-04 conditional novelty 6.0 of 10

    SideInfo-RoPE encodes reference-identity agreement as an extra rotary axis, disambiguating similar characters in multi-reference multi-shot video generation while keeping full semantic attention.

  3. 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

49 extracted references · 13 canonical work pages · cited by 3 Pith papers

  1. [1]

    https://app.klingai.com/cn/image-to-video/frame-mode/new

    Kling. https://app.klingai.com/cn/image-to-video/frame-mode/new

    work page

  2. [2]

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-vl technical report.arXiv preprint arXiv:2502.13923, 2025

    arXiv 2025

  3. [3]

    Stable video diffusion: Scaling latent video diffusion models to large datasets.arXiv preprint arXiv:2311.15127, 2023

    Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets.arXiv preprint arXiv:2311.15127, 2023

    arXiv 2023

  4. [4]

    Video generation models as world simulators, 2024

    Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators, 2024. URLhttps://openai.com/research/video-generation-models-as-world-simulators

    work page 2024

  5. [5]

    Goku: Flow based video generative foundation models.arXiv preprint arXiv:2502.04896, 2025

    Shoufa Chen, Chongjian Ge, Yuqi Zhang, Yida Zhang, Fengda Zhu, Hao Yang, Hongxiang Hao, Hui Wu, Zhichao Lai, Yifei Hu, et al. Goku: Flow based video generative foundation models.arXiv preprint arXiv:2502.04896, 2025

    arXiv 2025

  6. [6]

    Multi-subject open-set personalization in video generation

    Tsai-Shien Chen, Aliaksandr Siarohin, Willi Menapace, Yuwei Fang, Kwot Sin Lee, Ivan Skorokhodov, Kfir Aberman, Jun-Yan Zhu, Ming-Hsuan Yang, and Sergey Tulyakov. Multi-subject open-set personalization in video generation. arXiv preprint arXiv:2501.06187, 2025

    arXiv 2025

  7. [7]

    Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling.arXiv preprint arXiv:2412.05271, 2024

    Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, et al. Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling.arXiv preprint arXiv:2412.05271, 2024

    arXiv 2024

  8. [8]

    Arcface: Additive angular margin loss for deep face recognition

    Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690–4699, 2019

    2019

Show all 49 references
  1. [9]

    Cinema: Coherent multi-subject video generation via mllm-based guidance

    Yufan Deng, Xun Guo, Yizhi Wang, Jacob Zhiyuan Fang, Angtian Wang, Shenghai Yuan, Yiding Yang, Bo Liu, Haibin Huang, and Chongyang Ma. Cinema: Coherent multi-subject video generation via mllm-based guidance. arXiv preprint arXiv:2503.10391, 2025

    2025 arXiv

  2. [10]

    Skyreels-a2: Compose anything in video diffusion transformers.arXiv preprint arXiv:2504.02436, 2025

    Zhengcong Fei, Debang Li, Di Qiu, Jiahua Wang, Yikun Dou, Rui Wang, Jingtao Xu, Mingyuan Fan, Guibin Chen, Yang Li, et al. Skyreels-a2: Compose anything in video diffusion transformers.arXiv preprint arXiv:2504.02436, 2025

    2025 arXiv

  3. [11]

    I2vcontrol-camera: Precise video camera control with adjustable motion strength

    Wanquan Feng, Jiawei Liu, Pengqi Tu, Tianhao Qi, Mingzhen Sun, Tianxiang Ma, Songtao Zhao, Siyu Zhou, and Qian He. I2vcontrol-camera: Precise video camera control with adjustable motion strength. 2025

    2025

  4. [12]

    Animatediff: Animate your personalized text-to-image diffusion models without specific tuning

    Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. In 12th International Conference on Learning Representations, ICLR 2...

    2024

  5. [13]

    Pulid: Pure and lightning id customization via contrastive alignment.arXiv preprint arXiv:2404.16022, 2024

    Zinan Guo, Yanze Wu, Zhuowei Chen, Lang Chen, and Qian He. Pulid: Pure and lightning id customization via contrastive alignment.arXiv preprint arXiv:2404.16022, 2024

    2024 arXiv

  6. [14]

    Classifier-free diffusion guidance.arXiv preprint arXiv:2207.12598, 2022

    Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance.arXiv preprint arXiv:2207.12598, 2022

    2022 arXiv

  7. [15]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840–6851, 2020

    2020

  8. [16]

    Animate anyone: Consistent and controllable image-to-video synthesis for character animation

    Li Hu. Animate anyone: Consistent and controllable image-to-video synthesis for character animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8153–8163, 2024

    2024

  9. [17]

    Hunyuancustom: A multimodal-driven architecture for customized video generation.arXiv preprint arXiv:2505.04512, 2025

    Teng Hu, Zhentao Yu, Zhengguang Zhou, Sen Liang, Yuan Zhou, Qin Lin, and Qinglin Lu. Hunyuancustom: A multimodal-driven architecture for customized video generation.arXiv preprint arXiv:2505.04512, 2025

    2025 arXiv

  10. [18]

    Conceptmaster: Multi-concept video customization on diffusion transformer models without test-time tuning

    Yuzhou Huang, Ziyang Yuan, Quande Liu, Qiulin Wang, Xintao Wang, Ruimao Zhang, Pengfei Wan, Di Zhang, and Kun Gai. Conceptmaster: Multi-concept video customization on diffusion transformer models without test-time tuning. arXiv preprint arXiv:2501.04698, 2025

    2025 arXiv

  11. [19]

    Vbench: Comprehensive benchmark suite for video generative models

    Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. 12 In Proceedings of the IEEE/CVF Conference on Computer Vision and ...

    2024

  12. [20]

    Vace: All-in-one video creation and editing

    Zeyinzi Jiang, Zhen Han, Chaojie Mao, Jingfeng Zhang, Yulin Pan, and Yu Liu. Vace: All-in-one video creation and editing. arXiv preprint arXiv:2503.07598, 2025

    2025 arXiv

  13. [21]

    Fulldit: Multi-task video generative foundation model with full attention.arXiv preprint arXiv:2503.19907, 2025

    Xuan Ju, Weicai Ye, Quande Liu, Qiulin Wang, Xintao Wang, Pengfei Wan, Di Zhang, Kun Gai, and Qiang Xu. Fulldit: Multi-task video generative foundation model with full attention.arXiv preprint arXiv:2503.19907, 2025

    2025 arXiv

  14. [22]

    Videopoet: a large language model for zero-shot video generation

    Dan Kondratyuk, Lijun Yu, Xiuye Gu, José Lezama, Jonathan Huang, Grant Schindler, Rachel Hornung, Vighnesh Birodkar, Jimmy Yan, Ming-Chang Chiu, et al. Videopoet: a large language model for zero-shot video generation. In Proceedings of the 41st International Conference on Mach...

    2024

  15. [23]

    Hunyuanvideo: A systematic framework for large video generative models.arXiv preprint arXiv:2412.03603, 2024

    Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, et al. Hunyuanvideo: A systematic framework for large video generative models.arXiv preprint arXiv:2412.03603, 2024

    2024 arXiv

  16. [24]

    Movie weaver: Tuning-free multi-concept video personalization with anchored prompts

    Feng Liang, Haoyu Ma, Zecheng He, Tingbo Hou, Ji Hou, Kunpeng Li, Xiaoliang Dai, Felix Juefei-Xu, Samaneh Azadi, Animesh Sinha, et al. Movie weaver: Tuning-free multi-concept video personalization with anchored prompts. arXiv preprint arXiv:2502.07802, 2025

    2025 arXiv

  17. [25]

    Flow matching for generative modeling

    Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747, 2022

    2022 arXiv

  18. [26]

    Improving video generation with human feedback.arXiv preprint arXiv:2501.13918, 2025

    Jie Liu, Gongye Liu, Jiajun Liang, Ziyang Yuan, Xiaokun Liu, Mingwu Zheng, Xiele Wu, Qiulin Wang, Wenyu Qin, Menghan Xia, et al. Improving video generation with human feedback.arXiv preprint arXiv:2501.13918, 2025

    2025 arXiv

  19. [27]

    Phantom: Subject-consistent video generation via cross-modal alignment.arXiv preprint arXiv:2502.11079, 2025

    Lijie Liu, Tianxiang Ma, Bingchuan Li, Zhuowei Chen, Jiawei Liu, Qian He, and Xinglong Wu. Phantom: Subject-consistent video generation via cross-modal alignment.arXiv preprint arXiv:2502.11079, 2025

    2025 arXiv

  20. [28]

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

    Xingchao Liu, Chengyue Gong, et al. Flow straight and fast: Learning to generate and transfer data with rectified flow. In The Eleventh International Conference on Learning Representations

  21. [29]

    Dreamo: A unified framework for image customization.arXiv preprint arXiv:2504.16915, 2025

    Chong Mou, Yanze Wu, Wenxu Wu, Zinan Guo, Pengze Zhang, Yufeng Cheng, Yiming Luo, Fei Ding, Shiwen Zhang, Xinghui Li, et al. Dreamo: A unified framework for image customization.arXiv preprint arXiv:2504.16915, 2025

    2025

  22. [30]

    Dinov2: Learning robust visual features without supervision

    Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023

    2023 arXiv

  23. [31]

    Dinov2: Learning robust visual features without supervision

    Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. Transactions on Machine Learning Research Journal, 2024

    2024

  24. [32]

    Dreambench++: A human-aligned benchmark for personalized image generation.arXiv preprint arXiv:2406.16855, 2024

    Yuang Peng, Yuxin Cui, Haomiao Tang, Zekun Qi, Runpei Dong, Jing Bai, Chunrui Han, Zheng Ge, Xiangyu Zhang, and Shu-Tao Xia. Dreambench++: A human-aligned benchmark for personalized image generation.arXiv preprint arXiv:2406.16855, 2024

    2024 arXiv

  25. [33]

    Movie gen: A cast of media foundation models.arXiv preprint arXiv:2410.13720, 2024

    Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, et al. Movie gen: A cast of media foundation models.arXiv preprint arXiv:2410.13720, 2024

    2024 arXiv

  26. [34]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pa...

    2021

  27. [35]

    Magi-1: Autoregressive video generation at scale, 2025

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

    2025

  28. [36]

    Laion-5b: An open large-scale dataset for training next generation image-text models.Advances in neural information processing systems, 35:25278–25294, 2022

    Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models.Advances in neural informat...

    2022

  29. [37]

    Seaweed-7b: Cost-effective training of video generation foundation model.arXiv preprint arXiv:2504.08685, 2025

    Team Seawead, Ceyuan Yang, Zhijie Lin, Yang Zhao, Shanchuan Lin, Zhibei Ma, Haoyuan Guo, Hao Chen, Lu Qi, Sen Wang, et al. Seaweed-7b: Cost-effective training of video generation foundation model.arXiv preprint arXiv:2504.08685, 2025

    2025 arXiv

  30. [38]

    1st solution in google universal image embedding.https://www.kaggle.com/ datasets/louieshao/guieweights0732, april 2023

    Shihao Shao and Qinghua Cui. 1st solution in google universal image embedding.https://www.kaggle.com/ datasets/louieshao/guieweights0732, april 2023

    2023

  31. [40]

    Wan: Open and advanced large-scale video generative models.arXiv preprint arXiv:2503.20314, 2025

    Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianxiao Yang, Jianyuan Zeng, et al. Wan: Open and advanced large-scale video generative models.arXiv preprint arXiv:2503.20314, 2025

    2025 arXiv

  32. [41]

    Koala-36m: A large-scale video dataset improving consistency between fine-grained conditions and video content.arXiv preprint arXiv:2410.08260, 2024

    Qiuheng Wang, Yukai Shi, Jiarong Ou, Rui Chen, Ke Lin, Jiahao Wang, Boyuan Jiang, Haotian Yang, Mingwu Zheng, Xin 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

    arXiv

  33. [42]

    Magicvideo-v2: Multi-stage high-aesthetic video generation.arXiv preprint arXiv:2401.04468, 2024

    Weimin Wang, Jiawei Liu, Zhijie Lin, Jiangqiao Yan, Shuo Chen, Chetwin Low, Tuyen Hoang, Jie Wu, Jun Hao Liew, Hanshu Yan, et al. Magicvideo-v2: Multi-stage high-aesthetic video generation.arXiv preprint arXiv:2401.04468, 2024

    2024 arXiv

  34. [43]

    Objectmate: Arecurrencepriorforobjectinsertionandsubject-drivengeneration

    Daniel Winter, Asaf Shul, Matan Cohen, Dana Berman, Yael Pritch, Alex Rav-Acha, and Yedid Hoshen. Objectmate: Arecurrencepriorforobjectinsertionandsubject-drivengeneration. arXivpreprintarXiv:2412.08645, 2024

    2024 arXiv

  35. [44]

    Automated movie generation via multi-agent cot planning.arXiv preprint arXiv:2503.07314, 2025

    Weijia Wu, Zeyu Zhu, and Mike Zheng Shou. Automated movie generation via multi-agent cot planning.arXiv preprint arXiv:2503.07314, 2025

    2025 arXiv

  36. [45]

    Demystifying clip data

    Hu Xu, Saining Xie, Xiaoqing Tan, Po-Yao Huang, Russell Howes, Vasu Sharma, Shang-Wen Li, Gargi Ghosh, Luke Zettlemoyer, and Christoph Feichtenhofer. Demystifying clip data. InThe TwelfthInternational Conference on Learning Representations

  37. [46]

    Magicanimate: Temporally consistent human image animation using diffusion model

    Zhongcong Xu, Jianfeng Zhang, Jun Hao Liew, Hanshu Yan, Jia-Wei Liu, Chenxu Zhang, Jiashi Feng, and Mike Zheng Shou. Magicanimate: Temporally consistent human image animation using diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogn...

    2024

  38. [47]

    An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. Qwen2. 5 technical report.arXiv preprint arXiv:2412.15115, 2024

    2024 arXiv

  39. [48]

    Cogvideox: Text-to-video diffusion models with an expert transformer.arXiv preprint arXiv:2408.06072, 2024

    Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer.arXiv preprint arXiv:2408.06072, 2024

    2024 arXiv

  40. [49]

    Make pixels dance: High-dynamic video generation

    Yan Zeng, Guoqiang Wei, Jiani Zheng, Jiaxin Zou, Yang Wei, Yuchen Zhang, and Hang Li. Make pixels dance: High-dynamic video generation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8850–8860, 2024

    2024

  41. [50]

    Concat-id: Towards universal identity- preserving video synthesis.arXiv preprint arXiv:2503.14151, 2025

    Yong Zhong, Zhuoyi Yang, Jiayan Teng, Xiaotao Gu, and Chongxuan Li. Concat-id: Towards universal identity- preserving video synthesis.arXiv preprint arXiv:2503.14151, 2025. 14 Appendix A The Limitations of Synthetic Data We try two SOTA models, GPT4o and DreamO [29] to generat...

    2025 arXiv

Pith tools

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