Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

PolyVivid: Vivid Multi-Subject Video Generation with Cross-Modal Interaction and Enhancement

T0 review · 3 major / 5 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 PolyVivid claims that end-to-end multi-subject video customization can preserve each subject's identity and follow text-specified interactions, outperforming open-source and commercial baselines on identity fidelity, text alignment, and…

desk verdict A coherent, original architecture for multi-subject video customization, but the headline 'outperforms' claim is not yet statistically supported by the reported numbers. read the letter →

arxiv 2506.07848 v1 pith:OSXLWJWW submitted 2025-06-09 cs.CV cs.AI

Teng Hu , Zhentao Yu , Zhengguang Zhou , Jiangning Zhang , Yuan Zhou , Qinglin Lu , Ran Yi This is my paper
classification cs.CVcs.AI
keywords multi-subjectvideocustomizationidentitypreservationtext-imagefusion3D-RoPEcross-modalinteractiondiffusiontransformerMLLMdatapipelinegeneration
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

The paper seeks to establish that multi-subject video customization — generating a video from several reference images plus a text prompt describing how the subjects interact — can be done end-to-end, with no per-subject fine-tuning, while keeping every subject recognizable and the interaction faithful to the text. PolyVivid's answer to the correspondence problem (which image is which entity in the prompt) is to embed the subject images into the text space of a vision-language model through the template "The [entity] looks like ", then bind those text tokens to fine-grained VAE image tokens of the same subject using a shared 3D-RoPE grid. Identity-enhanced tokens are injected into every video frame through a cross-attention module built from the base model's own multimodal-attention weights, which the paper argues avoids the temporal identity decay of token concatenation and the misalignment of adapter-based injection. On a self-built benchmark of 100 image pairs, PolyVivid reports the best face and object identity similarity, the best video quality score (FVD), and the best or second-best text alignment among the open-source and commercial baselines it compares against. If the claim holds, arbitrary combinations of humans, animals, and objects could be placed into text-described scenes with consistent identity and no per-subject training.

What carries the argument

The load-bearing mechanism is the text-image interaction 3D-RoPE, a positional-binding scheme inside the text-image interaction module. All tokens live in one shared three-dimensional coordinate system: the prompt's <text> tokens keep their sequential order along the temporal axis at spatial position $(0,0)$; the LLaVA <image> tokens for subject $i$ are placed at the next temporal index with their genuine two-dimensional spatial layout; and the VAE image tokens $z_I$ of the same subject are placed one temporal index further, with spatially aligned indices. Because tokens belonging to the same subject land at nearly identical positions, the multimodal attention (MM-Attention) correlates them strongly: identity information flows from the VAE tokens into the LLaVA text tokens, and interaction semantics flow from the text into the image tokens. A second component, the attention-inherited identity injection, reparameterizes the pretrained MM-Attention's Key/Value matrices with LoRA to build a per-frame cross-attention from image tokens to video tokens (with a zero-initialized projection for training stability), which is what the paper credits for keeping identity consistent across frames.

What would settle it

Probe the fused embeddings before any video generation: encode a reference image through the LLaVA template branch and the VAE branch, extract the fused identity tokens, and measure their similarity to the original reference with the same face-sim and DINO-sim metrics used in Table 1. If the fused embeddings already lose identity, the reported identity gains cannot come from the claimed fusion pipeline. A companion observation: generate videos from pairs of visually similar subjects (two people who look alike, or two same-breed dogs) and check whether per-subject identity scores hold and whether the subjects swap roles; swapping would indicate the correspondence grounding failed.

Watch

Extended reading notes

Core claim

PolyVivid's central claim is that the combination of three mechanisms solves multi-subject customization where prior end-to-end methods fail: a LLaVA-based text-image fusion module that grounds each subject image to its textual entity via the structured template "The [entity] looks like <image>" separated by a <SEP> token; a 3D-RoPE-based enhancement module that lets identity flow from VAE-encoded image tokens into text tokens while interaction semantics flow back, producing identity-enhanced text tokens and interaction-enhanced image tokens; and an attention-inherited identity injection module that feeds the enhanced image tokens into the video token stream through cross-attention built from the pretrained multimodal-attention weights with low-rank adaptation, treating all frames equally. The paper argues this prevents both the subject-image/text confusion that plagues separate-condition models and the temporal identity drift that plagues token concatenation. On its 100-pair benchmark, PolyVivid reports the highest face similarity (Face-sim 0.642) and object similarity (DINO-sim 0.623), the lowest FVD (959.74), and the best and second-best CLIP text-alignment scores among VACE, SkyReels-A2, Keling, Vidu, and Pika.

Load-bearing premise

The pipeline assumes that LLaVA's text-space embeddings and the VAE's image tokens, after fusion, still retain enough fine-grained identity detail for the video model to reconstruct each subject, but the paper never directly measures identity retention inside those fused embeddings.

Editorial extensions

If this is right

  • Multi-subject video customization becomes a single-pass, zero-fine-tuning process: any new set of reference images can be inserted into the structured template and generated immediately.
  • Because subjects are laid out iteratively along the temporal axis of the 3D-RoPE grid, the same mechanism extends to three or more subjects, which the paper demonstrates qualitatively.
  • Per-frame cross-attention injection means identity guidance should not decay with video length, so longer videos should keep subject appearance consistent in a way prefix-concatenation methods cannot.
  • The data pipeline's clique-based subject consolidation offers a reusable recipe for building cleaner multi-subject training data for controllable video generation beyond this model.

Reading between the lines

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

  • My inference: the 3D-RoPE subject-binding idea is the most transferable piece and is not obviously tied to the HunyuanVideo backbone; any diffusion transformer with 3D-RoPE and multimodal attention could adopt the same binding, so its value should be testable in other backbones.
  • The paper's ablations toggle whole modules but never isolate the positional binding itself, so a cleaner experiment — shared versus deliberately unaligned 3D-RoPE positions with everything else fixed — would directly quantify how much of the identity gain the 3D-RoPE alignment contributes.
  • The benchmark is self-constructed (100 image pairs with prompts generated by QWen2.5-VL), so the ranking against commercial systems should be read as indicative until an independent benchmark reproduces it.
  • In practice, reference images with unusual lighting, occlusion, or low resolution would stress the VAE identity branch; robustness to such inputs is untested, and that is the first place to probe before relying on the method.
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

3 major / 5 minor

Summary. The paper proposes PolyVivid, a multi-subject video customization framework built on HunyuanVideo. The method comprises a VLLM-based text-image fusion module that grounds subject images to textual entities, a 3D-RoPE-based identity-interaction enhancement module for structured bidirectional fusion of text and image embeddings, an attention-inherited identity injection module that fuses identity features into the video generation process, and an MLLM-based data construction pipeline with clique-based subject consolidation. The evaluation compares PolyVivid against six baselines (three commercial APIs and three open-source methods) on a self-built 100-pair test set, reporting Face-sim, DINO-sim, CLIP-B/L, FVD, and temporal consistency, along with an ablation study. The central claim is that PolyVivid achieves superior performance in identity fidelity, video realism, and subject alignment, outperforming existing open-source and commercial baselines.

Significance. If the quantitative claims are substantiated, PolyVivid would be a meaningful advance in multi-subject video customization, particularly in its use of a VLLM to establish image-text grounding and its attention-inherited injection design that avoids temporal imbalance. The paper is clearly written, the architecture is well motivated, and the qualitative results suggest genuine capability in preserving multiple identities and generating text-specified interactions. However, the evidence for the headline 'outperforms' claim is currently statistically weak, and the evaluation protocol has several under-specified aspects. As presented, the significance of the result is conditional on additional uncertainty quantification and a more rigorous evaluation protocol.

major comments (3)
  1. [§5.2, Table 1] The central claim of outperforming all baselines is not statistically supported. Table 1 reports only point estimates with no standard deviations, confidence intervals, or significance tests, despite being computed from only 100 generated videos per method. The margins on CLIP-B (0.336 vs 0.335 for VACE and 0.336 for Vidu) and CLIP-L (0.281 vs 0.282 for Vidu, where PolyVivid is second-best) are at or below typical estimator noise. The FVD advantage over Pika (959.74 vs 980.49, a 20.7-point gap) is also small relative to the variance of FVD at n=100, particularly because FVD is computed against a generic 1,500-video reference set rather than a condition-matched reference. Without uncertainty quantification, the data do not support the abstract's claim of 'superior performance' over all baselines.
  2. [§5.2 and Appendix A.2] The evaluation protocol is under-specified, which affects the reproducibility and fairness of the comparison. The test set is self-built from 100 manually collected object images and 100 generated human images, randomly paired into 100 pairs with prompts produced by Qwen2.5-VL; there is no description of how commercial baselines (Vidu, Keling, Pika, Hailuo) were prompted, whether the same structured template or prompt format was used, whether multiple seeds or generation settings were attempted, or how API outputs were post-processed (e.g., resolution, duration, frame rate). Without this protocol detail, it is difficult to assess whether the reported margins reflect a real advantage of PolyVivid or artifacts of prompt/interface differences.
  3. [§5.3, Table 2] The ablation study also reports only point estimates and cannot establish that the proposed components contribute beyond noise. The incremental gains from adding the text-image interaction 3D-RoPE and the identity injection module (e.g., Face-sim from 0.601 to 0.642, DINO-sim from 0.605 to 0.623) are presented as evidence of effectiveness, but without variance or significance testing these could be within the noise of the 100-video evaluation. Moreover, the mechanism claimed in Sec. 4.2—that the fused embeddings zT,I and zI retain enough identity and correspondence information for downstream generation—is not directly validated by any diagnostic; the only evidence is the end-to-end Table 1, which has the statistical limitations noted above. A direct probing experiment (e.g., measuring identity similarity in the fused embeddings before injection) would strengthen the causal interpretation.
minor comments (5)
  1. [Figure 1 caption] Typo: 'waling' should be 'walking'.
  2. [Figure 8 caption] Typo: 'coffe' should be 'coffee'.
  3. [§4.2, first sentence] Unclosed parenthesis in 'The LLaVA model used in our text-image fusion (Sec. 4.1, as a multimodal understanding framework'; the phrase should be '(Sec. 4.1)' or restructured.
  4. [§5.2, Table 1] The metric 'Face-sim' is not defined in the main text; it is defined only in Appendix A.2. Consider defining all metrics where first used in the main paper.
  5. [Related Work and Sec. 1] The relationship to HunyuanCustom [16] is mentioned only in passing as a related work; since PolyVivid is built on the same base model and shares authors, an explicit sentence describing the difference (e.g., multi-subject support and the new fusion/injection modules) would help position the contribution.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: PolyVivid's architecture and ablations are self-contained, and its headline claim rests on external baselines rather than on a reduced or fitted identity.

full rationale

The derivation chain runs from Eq. (1)'s LLaVA template through the MM-attention interaction of Eq. (2), the 3D-RoPE indices of Eqs. (4)-(6), and the cross-attention injection of Eqs. (7)-(8). None of these equations defines a predicted quantity in terms of the measured identity scores; Face-sim, DINO-sim, CLIP-B/L, FVD, and temporal consistency are computed post hoc on generated videos against reference images and a fixed real-video set. The base model (HunyuanVideo) and the VLLM (LLaVA) are external pretrained systems cited as starting points, not as justifications that already contain the multi-subject customization result. The only co-authored references (ID-Animator, MotionMaster, HunyuanCustom) appear in literature or related-work context and are not used to justify the claimed superiority or to forbid alternative designs. The ablation study compares variants of the proposed modules rather than renaming a fitted parameter as a prediction. Statistical issues such as missing confidence intervals or the small FVD margin are evidence-quality concerns, not circularity. Accordingly, no circular step is identified.

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

The central claim is architectural and empirical, so the ledger contains no fitted physical law. The key unproved inputs are pretrained components: LLaVA grounding, HunyuanVideo VAE, and MM-Attention adaptation. Hand-chosen thresholds in the data pipeline shape training data but do not appear in a derived equation. No new physical entities are introduced.

free parameters (3)
  • MLLM segmentation CLIP validity threshold = not reported
    Hand-chosen threshold on CLIP score in Appendix A.3 to accept or reject segmented subjects; affects training data quality and indirectly the identity preservation claim.
  • Subject graph edge-distance threshold = not reported
    Threshold connecting subject images in the clique-based consolidation step; hand-chosen and not specified numerically.
  • Maximum-clique retention fraction = 1/3
    Subjects appearing in fewer than one-third of detected frames are removed from the training data; an ad hoc threshold shaping subject consistency.
assumptions (4)
  • domain assumption Pretrained HunyuanVideo MM-Attention can be adapted by LoRA to fuse text and image tokens without destroying pretrained video generation behavior.
    Relied on throughout Sec. 4.2 and 4.3; if the adapted attention fails to preserve the pretrained generation capability, the method would degrade.
  • domain assumption LLaVA's interleaved text-image embeddings capture the semantic correspondence between each subject image and its textual entity.
    Sec. 4.1 assumes the structured template of Eq. (1) lets LLaVA ground each image to the correct prompt entity.
  • domain assumption The VAE encoder of HunyuanVideo produces image tokens that retain fine-grained identity details beyond LLaVA semantics.
    Sec. 4.2 states LLaVA lacks fine details and relies on VAE zI to provide them; no direct measurement of identity retention in those tokens is given.
  • domain assumption Florence2 detection, SAM2 segmentation, and CLIP verification produce accurate subject masks for training data.
    Appendix A.3 describes the MLLM-based data pipeline; segmentation errors propagate into subject identity learning and interaction modeling.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PolyVivid: Vivid Multi-Subject Video Generation with Cross-Modal Interaction and Enhancement." pith.science (2026). https://pith.science/paper/OSXLWJWW

@misc{pith2026250607848,
  author       = {Pith},
  title        = {Pith review of: PolyVivid: Vivid Multi-Subject Video Generation with Cross-Modal Interaction and Enhancement},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OSXLWJWW}},
  note         = {Machine review of arXiv:2506.07848}
}
read the original abstract

Despite recent advances in video generation, existing models still lack fine-grained controllability, especially for multi-subject customization with consistent identity and interaction. In this paper, we propose PolyVivid, a multi-subject video customization framework that enables flexible and identity-consistent generation. To establish accurate correspondences between subject images and textual entities, we design a VLLM-based text-image fusion module that embeds visual identities into the textual space for precise grounding. To further enhance identity preservation and subject interaction, we propose a 3D-RoPE-based enhancement module that enables structured bidirectional fusion between text and image embeddings. Moreover, we develop an attention-inherited identity injection module to effectively inject fused identity features into the video generation process, mitigating identity drift. Finally, we construct an MLLM-based data pipeline that combines MLLM-based grounding, segmentation, and a clique-based subject consolidation strategy to produce high-quality multi-subject data, effectively enhancing subject distinction and reducing ambiguity in downstream video generation. Extensive experiments demonstrate that PolyVivid achieves superior performance in identity fidelity, video realism, and subject alignment, outperforming existing open-source and commercial baselines.

Figures

Figures reproduced from arXiv: 2506.07848 by the authors.

Figure 1
Figure 1. PolyVivid can generate high-quality customized videos from multiple subject images and a [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Framework of our PolyVivid: the text prompt and reference image are fused by the VLLM￾based text-image fusion module. Then, a 3D RoPE-based identity-interaction enhancement module is employed to enhance the text-image interaction. The enhanced image tokens are injected by an MM cross-attention module, which helps preserve the identities while ensuring good subject interaction. segmentation methods can better disting… view at source ↗
Figure 3
Figure 3. Comparison of the condition injection strategies for MM-DiT. For the ‘image’ tokens zI , we assign them with 3D RoPE of temporal index m1 + 2 and spatial indices aligned with those of the <image> tokens. This alignment facilitates efficient pixel-by-pixel interaction between the two sets of image tokens. The RoPE position indices for the first subject’s ‘image’ tokens zI are set as: IdxRoP E(z i I ) = (m1 + 2, ⌊ i h… view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Comparison on multi-subject video customization. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Examples of the test set, which contains images from diverse categories, such as human, [PITH_FULL_IMAGE:figures/full_fig_p015_5.png]
Figure 6
Figure 6. Figure 6: Comparison on human-object customization. [PITH_FULL_IMAGE:figures/full_fig_p016_6.png]
Figure 7
Figure 7. Figure 7: Comparison on human-human and animal-animal customizations. [PITH_FULL_IMAGE:figures/full_fig_p016_7.png]
Figure 8
Figure 8. Figure 8: Comparison on three-subject customization. [PITH_FULL_IMAGE:figures/full_fig_p017_8.png]
Figure 9
Figure 9. Figure 9: More results on multi-subject customization. [PITH_FULL_IMAGE:figures/full_fig_p018_9.png]
Figure 10
Figure 10. Figure 10: More results on three-subject customization. [PITH_FULL_IMAGE:figures/full_fig_p019_10.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Aura: Consistent Multi-Subject Video Generation via VLM-Grounded Semantic Alignment

    cs.CV 2026-07 conditional novelty 6.0 of 10

    Aura combines VLM meta-queries, T5-teacher alignment, subject-aware RoPE shifts, memory tokens, and a large AIGC-curated dataset to claim SOTA multi-element subject-to-video generation under OpenS2V-Eval Total score.

Reference graph

Works this paper leans on

48 extracted references · 15 canonical work pages · cited by 1 Pith paper

  1. [1]

    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

    arXiv 2025

  2. [2]

    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

    arXiv 2023

  3. [3]

    Carreira and A

    J. Carreira and A. Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. Inproceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299–6308, 2017

    work page 2017

  4. [4]

    Chefer, S

    H. Chefer, S. Zada, R. Paiss, A. Ephrat, O. Tov, M. Rubinstein, L. Wolf, T. Dekel, T. Michaeli, and I. Mosseri. Still-moving: Customized video generation without customized video data. ACM Transactions on Graphics (TOG), 43(6):1–11, 2024

    work page 2024

  5. [5]

    H. Chen, X. Wang, Y . Zhang, Y . Zhou, Z. Zhang, S. Tang, and W. Zhu. Disenstudio: Customized multi-subject text-to-video generation with disentangled spatial control. InProceedings of the 32nd ACM International Conference on Multimedia, pages 3637–3646, 2024

    work page 2024

  6. [6]

    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. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13320–13331, 2024

    work page 2024

  7. [7]

    T.-S. Chen, A. Siarohin, W. Menapace, Y . Fang, K. S. Lee, I. Skorokhodov, K. Aberman, J.-Y . Zhu, M.-H. Yang, and S. Tulyakov. Multi-subject open-set personalization in video generation. arXiv preprint arXiv:2501.06187, 2025

    arXiv 2025

  8. [8]

    J. Deng, J. Guo, N. Xue, and S. 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

    work page 2019

Show all 48 references
  1. [9]

    Esser, S

    P. Esser, S. Kulal, A. Blattmann, R. Entezari, J. Müller, H. Saini, Y . Levi, D. Lorenz, A. Sauer, F. Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In Forty-first international conference on machine learning, 2024. 10

    2024

  2. [10]

    Z. Fei, D. Li, D. Qiu, J. Wang, Y . Dou, R. Wang, J. Xu, M. Fan, G. Chen, Y . Li, et al. Skyreels-a2: Compose anything in video diffusion transformers.arXiv preprint arXiv:2504.02436, 2025

    2025 arXiv

  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. arXiv preprint arXiv:2307.04725, 2023

    2023 arXiv

  4. [12]

    Hailuo.https://hailuoai.video/, 2025

    Hailuo. Hailuo.https://hailuoai.video/, 2025

    2025

  5. [13]

    X. He, Q. Liu, S. Qian, X. Wang, T. Hu, K. Cao, K. Yan, and J. Zhang. Id-animator: Zero-shot identity-preserving human video generation.arXiv preprint arXiv:2404.15275, 2024

    2024 arXiv

  6. [14]

    E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, W. Chen, et al. Lora: Low-rank adaptation of large language models.ICLR, 1(2):3, 2022

    2022

  7. [15]

    T. Hu, J. Zhang, R. Yi, Y . Wang, H. Huang, J. Weng, Y . Wang, and L. Ma. Motionmaster: Training-free camera motion transfer for video generation.arXiv preprint arXiv:2404.15789, 2024

    2024 arXiv

  8. [16]

    T. Hu, Z. Yu, Z. Zhou, S. Liang, Y . Zhou, Q. Lin, and Q. Lu. Hunyuancustom: A multimodal- driven architecture for customized video generation.arXiv preprint arXiv:2505.04512, 2025

    2025 arXiv

  9. [17]

    Huang, Z

    Y . Huang, Z. Yuan, Q. Liu, Q. Wang, X. Wang, R. Zhang, P. Wan, D. Zhang, and K. Gai. Conceptmaster: Multi-concept video customization on diffusion transformer models without test-time tuning.arXiv preprint arXiv:2501.04698, 2025

    2025 arXiv

  10. [18]

    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. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807–21818, 2024

    2024

  11. [19]

    Jiang, T

    Y . Jiang, T. Wu, S. Yang, C. Si, D. Lin, Y . Qiao, C. C. Loy, and Z. Liu. Videobooth: Diffusion- based video generation with image prompts. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6689–6700, 2024

    2024

  12. [20]

    Jiang, Z

    Z. Jiang, Z. Han, C. Mao, J. Zhang, Y . Pan, and Y . Liu. Vace: All-in-one video creation and editing.arXiv preprint arXiv:2503.07598, 2025

    2025 arXiv

  13. [21]

    Keling.https://klingai.com/cn/, 2025

    Keling. Keling.https://klingai.com/cn/, 2025

    2025

  14. [22]

    Khanam and M

    R. Khanam and M. Hussain. Yolov11: An overview of the key architectural enhancements. arXiv preprint arXiv:2410.17725, 2024

    2024 arXiv

  15. [23]

    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

    2024 arXiv

  16. [24]

    X. Lai, Z. Tian, Y . Chen, Y . Li, Y . Yuan, S. Liu, and J. Jia. Lisa: Reasoning segmentation via large language model. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9579–9589, 2024

    2024

  17. [25]

    Liang, Z

    S. Liang, Z. Yu, Z. Zhou, T. Hu, H. Wang, Y . Chen, Q. Lin, Y . Zhou, X. Li, Q. Lu, and Z. Chen. Omniv2v: Versatile video generation and editing via dynamic content manipulation, 2025. URL https://arxiv.org/abs/2506.01801

    2025 arXiv

  18. [26]

    H. Liu, C. Li, Q. Wu, and Y . J. Lee. Visual instruction tuning.Advances in neural information processing systems, 36:34892–34916, 2023

    2023

  19. [27]

    L. Liu, T. Ma, B. Li, Z. Chen, J. Liu, Q. He, and X. Wu. Phantom: Subject-consistent video generation via cross-modal alignment.arXiv preprint arXiv:2502.11079, 2025

    2025 arXiv

  20. [28]

    S. Liu, Z. Zeng, T. Ren, F. Li, H. Zhang, J. Yang, Q. Jiang, C. Li, J. Yang, H. Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. InEuropean Conference on Computer Vision, pages 38–55. Springer, 2024. 11

    2024

  21. [29]

    Y . Liu, K. Zhang, Y . Li, Z. Yan, C. Gao, R. Chen, Z. Yuan, Y . Huang, H. Sun, J. Gao, et al. Sora: A review on background, technology, limitations, and opportunities of large vision models. arXiv preprint arXiv:2402.17177, 2024

    2024 arXiv

  22. [30]

    Oquab, T

    M. Oquab, T. Darcet, T. Moutakanni, H. V o, M. Szafraniec, V . Khalidov, P. Fernandez, D. Haziza, F. Massa, A. El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023

    2023 arXiv

  23. [31]

    Peebles and S

    W. Peebles and S. Xie. Scalable diffusion models with transformers. InProceedings of the IEEE/CVF international conference on computer vision, pages 4195–4205, 2023

    2023

  24. [32]

    Pika.https://pika.art/, 2025

    Pika. Pika.https://pika.art/, 2025

    2025

  25. [33]

    Polyak, A

    A. Polyak, A. Zohar, A. Brown, A. Tjandra, A. Sinha, A. Lee, A. Vyas, B. Shi, C.-Y . Ma, C.-Y . Chuang, D. Yan, D. Choudhary, D. Wang, G. Sethi, G. Pang, H. Ma, et al. Movie gen: A cast of media foundation models, 2025. URLhttps://arxiv.org/abs/2410.13720

    2025 arXiv

  26. [34]

    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. InInternational conference on machine learning, pages 8748–8763. PmLR, 2021

    2021

  27. [35]

    N. Ravi, V . Gabeur, Y .-T. Hu, R. Hu, C. Ryali, T. Ma, H. Khedr, R. Rädle, C. Rolland, L. Gustafson, et al. Sam 2: Segment anything in images and videos.arXiv preprint arXiv:2408.00714, 2024

    2024 arXiv

  28. [36]

    Rombach, A

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

    2022

  29. [37]

    N. Ruiz, Y . Li, V . Jampani, Y . Pritch, M. Rubinstein, and K. Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500–22510, 2023

    2023

  30. [38]

    Z. Tan, S. Liu, X. Yang, Q. Xue, and X. Wang. Ominicontrol: Minimal and universal control for diffusion transformer.arXiv preprint arXiv:2411.15098, 3, 2024

    2024 arXiv

  31. [39]

    Vidu.https://www.vidu.cn/, 2025

    Vidu. Vidu.https://www.vidu.cn/, 2025

    2025

  32. [40]

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

    2025 arXiv

  33. [41]

    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

    arXiv

  34. [42]

    Z. Wang, A. Li, L. Zhu, Y . Guo, Q. Dou, and Z. Li. Customvideo: Customizing text-to-video generation with multiple subjects.arXiv preprint arXiv:2401.09962, 2024

    2024

  35. [43]

    T. Wu, Y . Zhang, X. Wang, X. Zhou, G. Zheng, Z. Qi, Y . Shan, and X. Li. Customcrafter: Customized video generation with preserving motion and concept composition abilities. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 8469–8477, 2025

    2025

  36. [44]

    B. Xiao, H. Wu, W. Xu, X. Dai, H. Hu, Y . Lu, M. Zeng, C. Liu, and L. Yuan. Florence-2: Advancing a unified representation for a variety of vision tasks. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4818–4829, 2024

    2024

  37. [45]

    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

    2024 arXiv

  38. [46]

    H. Ye, J. Zhang, S. Liu, X. Han, and W. Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models.arXiv preprint arXiv:2308.06721, 2023. 12

    2023 arXiv

  39. [47]

    S. Yuan, J. Huang, X. He, Y . Ge, Y . Shi, L. Chen, J. Luo, and L. Yuan. Identity-preserving text-to-video generation by frequency decomposition.arXiv preprint arXiv:2411.17440, 2024

    2024 arXiv

  40. [48]

    Y . Zhou, Q. Wang, Y . Cai, and H. Yang. Allegro: Open the black box of commercial-level video generation model.arXiv preprint arXiv:2410.15458, 2024. 13 A Appendix A.1 Overview In this supplementary material, we offer further details on implementation, present additional expe...

    2024 arXiv

Pith tools

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