Pith. sign in

REVIEW 4 major objections 5 minor 43 references

Identity-preserving text-to-video generation can be improved with no training at all: rewrite the prompt to include facial details, edit the reference image to match the prompt, and add a guidance term driven by the difference between a str

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

A training-free prompt, image, and guidance enhancement framework improves face consistency and video quality for identity-preserving text-to-video generation, winning the ACM Multimedia 2025 IPVG challenge.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection A useful training-free IPT2V recipe with real gains that get partly confounded by the MoE oracle selection; worth a serious referee, but the weak-model guidance and the MoE ablation need to be tightened. the 4 major comments →

arxiv 2509.01362 v1 pith:ZTLHYY4W submitted 2025-09-01 cs.CV cs.MM

Identity-Preserving Text-to-Video Generation via Training-Free Prompt, Image, and Guidance Enhancement

classification cs.CV cs.MM
keywords identity-preserving text-to-video generationtraining-free inferenceface-aware prompt rewritingreference image editingdiffusion guidanceidentity consistencyvideo qualitymixture-of-experts selection
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

Identity-preserving text-to-video generation usually forces a trade-off: either per-identity fine-tuning or large offline training datasets, and even then the generated video may lose the face or freeze the subject. This paper proposes TPIGE, a fully training-free alternative that claims to improve both identity preservation and video quality by upgrading the inputs before generation and steering the sampling after. It rewrites the text prompt with facial attributes extracted from the reference image; it re-generates the reference image so its appearance already contains prompt-relevant clothing or expression; and it adds an identity-aware guidance term to the classifier-free guidance update. On a 1,000-video test set, the method reports the best face-consistency scores among the compared methods, a first-place finish in its challenge, and higher preference in a 3,000-pair user study. The same enhancement strategies transfer to a second video model, which the paper offers as evidence of generality.

Core claim

The central claim is that the semantic gap between the reference image and the text prompt is the root cause of identity loss and frozen motion, and that it can be closed without training. Face Aware Prompt Enhancement makes the prompt identity-aware by inserting a short facial description from the reference image; Prompt Aware Reference Image Enhancement makes the image prompt-aware by regenerating it with the role- or action-specific appearance called for by the prompt. The third piece, ID-Aware Spatiotemporal Guidance Enhancement, modifies the denoising update to include the difference between the normal model prediction and a deliberately weakened, identity-agnostic prediction, treating

What carries the argument

ID-Aware Spatiotemporal Guidance Enhancement is the component that carries the paper's strongest claim: at each sampling step the update becomes the usual classifier-free guidance plus a weight times the difference between the noise predicted by the identity-conditioned model and the noise predicted by a 'weak' variant of the same model, formed by removing identity inputs and skipping selected layers. The paper asserts this difference estimates the score toward the reference identity and a high-quality video, so the guidance is doing in one term what separate identity and quality losses would do. The two pre-generation modules, face-aware prompt rewriting and prompt-aware reference image edi

Load-bearing premise

The load-bearing premise is that a weakened version of the video model, made by removing identity inputs and skipping some layers, serves as a valid low-quality baseline, so the difference between the normal and weak predictions points the generation toward the reference identity and high video quality; the paper gives no construction or derivation for that baseline, and if the approximation is wrong the unified guidance has no foundation.

What would settle it

Vary the choice of skipped layers used to build the weak model while keeping all other settings fixed. If the identity-consistency and quality gains remain roughly constant across many layer subsets, the guidance term is a robust score estimate; if a single arbitrary selection produces the reported gains and others do not, the method's guiding mechanism is not established.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Users can adapt video generation to a new face with only API calls and a few inference-time additions, no per-identity fine-tuning or paired video data.
  • The enhancements are model-agnostic: the same prompt-plus-guidance recipe improved a second open backbone, so the gains should stack onto newer video models as they appear.
  • The guidance term imports a video-quality objective directly into sampling, so results target identity and perceptual quality simultaneously rather than optimizing text alignment alone.
  • Because the mixture-of-experts selector picks per-sample among method variants, final outputs can mix strengths of different generation strategies without retraining.
  • The small drop in text-alignment score is presented as a deliberate trade-off: jointly conditioning on image and text, rather than text alone, balances prompt fidelity against identity consistency.

Where Pith is reading between the lines

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

  • The weak-model construction is currently underspecified: which layers get skipped is not stated and no derivation connects the degraded prediction to the imaginary low-quality baseline. A natural test is to replace the skipped layers with a random subset or a text-only run and see whether the identity and quality gain is preserved, which would distinguish the score-difference mechanism from the sp
  • Because the identity metrics score similarity to the original reference image while the video model receives the edited reference image, the image-enhancement module can help text alignment yet appear to hurt identity consistency by inspection; an evaluation that also measures consistency to the edited reference would isolate the real effect.
  • The mixture-of-experts result is obtained by generating six videos per sample and picking by a weighted score, so the reported gains come with roughly six-fold inference cost; a practical deployment would need to study that cost-quality trade-off or learn to route without full generation.
  • A stronger claim implicit in the paper is that identity and quality can share one gradient direction; if true, the same unified-guidance idea may apply to other personalization tasks, such as preserving a product logo or a character design across shots.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes TPIGE, a training-free framework for identity-preserving text-to-video generation built on the VACE backbone. It consists of three enhancements: (1) Face Aware Prompt Enhancement, which uses GPT-4o to insert facial descriptors into the text prompt; (2) Prompt Aware Reference Image Enhancement, which uses an identity-preserving image generator to edit the reference image so that it better matches the prompt; and (3) ID-Aware Spatiotemporal Guidance Enhancement, which adds a guidance term to the diffusion sampling step that is computed as the difference between a normal model prediction and a weakened, identity-agnostic prediction. A Mixture-of-Experts (MoE) strategy selects, per sample, the video with the highest weighted Overall Score among six generation methods (VACE, Phantom, Hailuo, PE, PE&IE, PE&GE). The authors report first place in the ACM MM 2025 IPVG challenge, ablation results on 50 samples, a user study, and a generalizability experiment on Phantom.

Significance. If the claims hold, TPIGE would be a useful low-cost alternative to fine-tuning-based identity-preserving video generation, and the external challenge result plus user study provide some real-world validation. The paper also shows an interesting direction: using a weakened model to define identity/quality guidance. However, the significance is substantially tempered by the MoE selection procedure, which optimizes the same metrics used for evaluation, and by the underspecified and underived GE formulation. The lack of statistical testing on the 50-sample evaluation further weakens the quantitative claims. The strengths are the training-free design, the attempt to attack the semantic gap from both the prompt and image sides, and the public code link, though the code availability does not compensate for missing weights and layer-skip details.

major comments (4)
  1. [§3.5 and Table 3] The MoE evaluation is circular. The Overall Score in §3.5 is defined as a weighted sum of GMEScore, CurScore, ArcScore, Motion Smoothness, and Imaging Quality, and the MoE strategy selects, for each sample, the video with the highest Overall Score among six methods. Reporting in Table 3 that the MoE row attains the best CurScore (0.5176), ArcScore (0.5007), and Overall Score (0.6337) is a monotonicity consequence of the selection rule, not evidence of generative improvement. The challenge submission in Table 1 uses this MoE selection, so the first-place result compares an oracle-selected output against other teams' single outputs; no no-MoE submission or equal best-of-N baseline is provided. Please report the selection weights w_i, evaluate a fixed single-method submission on the official protocol, and either use metrics not involved in selection or provide a direct per-method selected-v
  2. [§3.4, Eqs. (1)-(3)] The ID-Aware Spatiotemporal Guidance Enhancement is underdefined and lacks a derivation. Eq. (1) introduces an 'imaginary high-quality video y_g' and a 'low-quality label y_b' that are never defined; Eq. (3) uses a 'weak' model obtained by 'removing identity inputs and skipping selected layers,' but the skipped layer set is not specified. No argument is given for why this ablated network approximates the low-quality baseline log p(x_t|T_c,∅,y_b). The guidance term therefore has no established foundation. Additionally, the guidance weights W_c and W_i in Eqs. (2)-(3) are not reported anywhere, making the main enhancement irreproducible. Please define y_g and y_b (or reformulate the gradient without them), specify the exact weak-model construction, and report W_c and W_i for all experiments.
  3. [§4.2 and §4.3] All quantitative comparisons are point estimates on 50 samples with no variance, confidence intervals, or significance tests. The text in §4.3 claims the method 'significantly outperforms' other models on identity metrics, but with n=50 and one prompt per identity, sampling noise is nontrivial. This is especially important for the GE gains, which are the central contribution. Please provide paired statistical tests (e.g., bootstrap CIs or Wilcoxon signed-rank) on the Table 2 and Table 3 comparisons, or at a minimum per-sample score distributions.
  4. [§4.5, Table 3] The standalone effect of IE appears negative on the paper's own objective: +PE & IE decreases CurScore (0.4040→0.3510), ArcScore (0.3871→0.3425), and Overall Score (0.5815→0.5655) relative to +PE. The authors attribute this to the identity metric being computed against the original reference image, which is plausible, but it means IE's only reported benefit is through the MoE selection. Since MoE selection uses the same metrics, the claim that 'our MoE strategy also selects some videos generated by PE & IE' does not by itself demonstrate that IE improves the final output. Please provide a comparison of final outputs with and without IE under identical selection, or an analysis of what IE contributes that the metrics do not capture.
minor comments (5)
  1. [Throughout] Typos: 'A ware' appears in headings (e.g., 'Face A ware Prompt Enhancement'); 'ID ref c' notation is inconsistent. Please proofread.
  2. [Tables 2 and 3] The relationship between 'Ours' in Table 2 and '+PE & GE' in Table 3 is unclear; they have identical numeric values but this is not stated. Clarify whether Table 2 is the PE&GE configuration without MoE.
  3. [§3.1] The term 'training-free' is used loosely: the framework relies on GPT-4o and an identity-preserving image generator (ACE++), both of which are trained models. Clarify that 'training-free' refers to no additional training or fine-tuning of the video generation model.
  4. [§4.1] The evaluation metric description says FID is used to 'assess feature differences in the face regions,' but FID is typically a distribution-level metric on full images. Specify the exact implementation and region of interest.
  5. [§3.5] The MoE weights w_i are never disclosed, so the selection rule is not reproducible even with the code link. This is mentioned in the major comments, but it should also be stated explicitly in the paper.

Circularity Check

1 steps flagged

MoE selection is an oracle over the evaluation metrics themselves, so its reported gains and the first-place result are partly circular; PE/GE have independent support.

specific steps
  1. fitted input called prediction [Section 4.5, 'Effectiveness of the MoE Strategy' (with Section 3.5 and Table 3)]
    "We adopt the MoE strategy to select the video with the highest overall score for each sample from the results generated by different methods. We then compute the mean value of each metric (excluding CLIPScore and FID, as they are not involved in the calculation of the overall score) for the resulting set of videos. ... As shown in the last row of Table 3, almost all metrics reach the best performance, confirming the effectiveness of the MoE strategy for final challenge results."

    The Overall Score used for selection (Section 3.5) is a weighted sum of GMEScore, CurScore, ArcScore, Motion Smoothness, and Imaging Quality. Table 3 evaluates MoE on essentially the same metrics, and Section 4.1 says Face-Cur/Face-Arc 'correspond to CurScore and ArcScore'. Selecting the per-sample maximum of a weighted sum guarantees that the selected set's mean on those metrics cannot be lower than any single method's mean; the reported MoE gains (CurScore 0.4533->0.5176, ArcScore 0.4358->0.5007, Overall 0.5997->0.6337) are thus a mathematical consequence of the selection rule, not evidence that the generative enhancements improve identity/quality. The challenge submission used this MoE selection (Section 4.6), so the first-place quantitative result and the user study (MoE outputs vs sin

full rationale

The core PE/IE/GE contributions are not circular: Table 2 and the non-MoE rows of Table 3 show PE&GE alone improve CurScore/ArcScore/Overall over VACE (0.3105->0.4533 CurScore; 0.5488->0.5997 Overall) without any per-sample oracle selection, so those enhancements have independent empirical content. The substantial circularity is confined to the MoE strategy: it selects outputs by a weighted sum that includes the identity and quality metrics later used for evaluation and for the challenge ranking. Section 4.5 presents the resulting metric increases as 'confirming the effectiveness of the MoE strategy,' which is true by construction rather than by generative improvement. The GE formulation in Section 3.4 is under-specified (y_g and y_b are undefined, skipped layers are unspecified) and is better treated as an ansatz/rigor concern than as circularity. Overall, one central 'prediction'—that MoE improves results and that the first-place submission outperforms prior work—reduces to selection on the evaluation metric, so the paper deserves a 6 rather than a 0, but not a higher score because the training-free enhancements themselves stand on non-MoE comparisons.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 1 invented entities

The framework rests on several externally trained models and on an unvalidated weak-model heuristic. The only fully internal contribution that is independently confirmed by the user study is the PE+GE combination, while the MoE part introduces free weights and metric-based selection.

free parameters (4)
  • W_c (CFG guidance weight) = not reported
    Scales the standard text guidance term in Eq. 3; tuned by hand or validation, but no value or sensitivity analysis is given.
  • W_i (ID-aware guidance weight) = not reported
    Scales the identity and quality guidance term in Eq. 3; tuned, with no reported value or sensitivity study.
  • MoE metric weights w_i = not reported
    Weights for GMEScore, CurScore, ArcScore, Motion Smoothness, and Imaging Quality in Section 3.5; no values are given despite being load-bearing for the final challenge submission.
  • Weak-model layer skip set = not specified
    Which layers of the video model are skipped to form the weak prediction in Section 3.4; a critical design choice with no detail.
axioms (5)
  • standard math The denoising network epsilon_theta approximates the score of the data distribution, so differences between noise predictions can be treated as gradients of log-density.
    This is the standard classifier-free guidance assumption invoked in Section 3.4.
  • domain assumption GPT-4o can extract facial attributes from a reference image and insert them into a prompt while preserving the original wording and meaning.
    Section 3.2 depends entirely on this for prompt enhancement.
  • domain assumption ACE++ preserves the subject's identity while editing the reference image to include prompt-specified attire or attributes.
    Section 3.3 depends on this for reference-image enhancement.
  • ad hoc to paper Removing identity inputs and skipping selected layers of the video model yields a weaker, identity-agnostic model whose prediction represents a low-quality baseline.
    Section 3.4, Eq. 3; this is an unvalidated modeling choice with no derivation or layer specification.
  • ad hoc to paper A weighted sum of GMEScore, CurScore, ArcScore, Motion Smoothness, and Imaging Quality is a reliable proxy for the best video.
    Section 3.5 uses this to select final outputs; weights are undisclosed and the proxy is validated only indirectly through challenge results.
invented entities (1)
  • imaginary high-quality video y_g and low-quality label y_b no independent evidence
    purpose: To formalize the score-gradient expression in Eq. 1 for identity and quality guidance.
    These conditioning labels are introduced in Section 3.4 but never defined, instantiated, or used directly in the final update; they justify the weak-model difference but carry no falsifiable prediction.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Identity-Preserving Text-to-Video Generation via Training-Free Prompt, Image, and Guidance Enhancement." pith.science (2026). https://pith.science/paper/ZTLHYY4W

@misc{pith2026250901362,
  author       = {Pith},
  title        = {Pith review of: Identity-Preserving Text-to-Video Generation via Training-Free Prompt, Image, and Guidance Enhancement},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZTLHYY4W}},
  note         = {Machine review of arXiv:2509.01362}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Identity-preserving text-to-video (IPT2V) generation creates videos faithful to both a reference subject image and a text prompt. While fine-tuning large pretrained video diffusion models on ID-matched data achieves state-of-the-art results on IPT2V, data scarcity and high tuning costs hinder broader improvement. We thus introduce a Training-Free Prompt, Image, and Guidance Enhancement (TPIGE) framework that bridges the semantic gap between the video description and the reference image and design sampling guidance that enhances identity preservation and video quality, achieving performance gains at minimal cost.Specifically, we first propose Face Aware Prompt Enhancement, using GPT-4o to enhance the text prompt with facial details derived from the reference image. We then propose Prompt Aware Reference Image Enhancement, leveraging an identity-preserving image generator to refine the reference image, rectifying conflicts with the text prompt. The above mutual refinement significantly improves input quality before video generation. Finally, we propose ID-Aware Spatiotemporal Guidance Enhancement, utilizing unified gradients to optimize identity preservation and video quality jointly during generation.Our method outperforms prior work and is validated by automatic and human evaluations on a 1000 video test set, winning first place in the ACM Multimedia 2025 Identity-Preserving Video Generation Challenge, demonstrating state-of-the-art performance and strong generality. The code is available at https://github.com/Andyplus1/IPT2V.git.

Figures

Figures reproduced from arXiv: 2509.01362 by Changcheng Hua, Jiayi Gao, Qingchao Chen, Yang Liu, Yuxin Peng.

Figure 1
Figure 1. Figure 1: Our method produces higher-quality identity-preserving videos than baselines: faces are clearer and match the [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Pipeline of TPIGE. TPIGE consists of three parts: (1) Face Aware Prompt Enhancement enriches the prompt with facial attributes; (2) Prompt Aware Reference Image Enhancement edits the reference image to incorporate prompt-aligned appearance cues; and (3) ID-Aware Spatiotemporal Guidance Enhancement guide the sampling for improved identity preservation and video quality. require collecting large amounts of d… view at source ↗
Figure 3
Figure 3. Figure 3: Visualization results of different methods [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

43 extracted references · 22 canonical work pages · 1 internal anchor

  1. [1]

    Fan Bao, Chendong Xiang, Gang Yue, Guande He, Hongzhou Zhu, Kaiwen Zheng, Min Zhao, Shilong Liu, Yaole Wang, and Jun Zhu. 2024. Vidu: a highly consistent, dynamic and skilled text-to-video generator with diffusion models. arXiv:2405.04233 [cs.CV] https://arxiv.org/abs/2405.04233

  2. [2]

    Hila Chefer, Shiran Zada, Roni Paiss, Ariel Ephrat, Omer Tov, Michael Rubinstein, Lior Wolf, Tali Dekel, Tomer Michaeli, and Inbar Mosseri. 2024. Still-moving: Customized video generation without customized video data. ACM Transactions on Graphics (TOG) 43, 6 (2024), 1–11

  3. [3]

    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. 2025. Multi-subject open-set personalization in video generation. In Proceedings of the Computer Vision and Pattern Recognition Conference. 6099–6110

  4. [4]

    Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. 2019. Arcface: Additive angular margin loss for deep face recognition. In CVPR. 4690–4699

  5. [5]

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. 2024. Scaling rectified flow transformers for high-resolution image synthesis. In Forty- first international conference on machine learning

  6. [6]

    Jiayi Gao, Zijin Yin, Changcheng Hua, Yuxin Peng, Kongming Liang, Zhanyu Ma, Jun Guo, and Yang Liu. 2025. Conmo: Controllable motion disentanglement and recomposition for zero-shot motion transfer. In Proceedings of the Computer Vision and Pattern Recognition Conference . 7191–7200

  7. [7]

    Xuanhua He, Quande Liu, Shengju Qian, Xin Wang, Tao Hu, Ke Cao, Keyu Yan, Man Zhou, and Jie Zhang. 2024. ID-Animator: Zero-Shot Identity-Preserving Human Video Generation. arXiv:2404.15275 [cs.CV] https://arxiv.org/abs/2404. 15275

  8. [8]

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. 2017. GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium. NeurIPS (Jan 2017)

  9. [9]

    Jonathan Ho and Tim Salimans. 2022. Classifier-free diffusion guidance. arXiv:2207.12598 [cs.CV] https://arxiv.org/abs/2207.12598

  10. [10]

    Yuge Huang, Yuhan Wang, Ying Tai, Xiaoming Liu, Pengcheng Shen, Shaoxin Li, Jilin Li, and Feiyue Huang. 2020. Curricularface: adaptive curriculum learning loss for deep face recognition. In CVPR. 5901–5910

  11. [11]

    Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuan- han Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. 2024. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 21807– 21818

  12. [12]

    Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. 2024. Gpt-4o system card. arXiv:2410.21276 [cs.CL] https://arxiv.org/abs/2410.21276

  13. [13]

    Zeyinzi Jiang, Zhen Han, Chaojie Mao, Jingfeng Zhang, Yulin Pan, and Yu Liu

  14. [14]

    Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, et al. 2024. Hunyuanvideo: A systematic framework for large video generative models. arXiv:2412.03603 [cs.CV] https: //arxiv.org/abs/2412.03603

  15. [15]

    Sizhe Li, Yiming Qin, Minghang Zheng, Xin Jin, and Yang Liu. 2024. Diff-bgm: A diffusion model for video background music generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 27348–27357

  16. [16]

    Zhen Li, Mingdeng Cao, Xintao Wang, Zhongang Qi, Ming-Ming Cheng, and Ying Shan. 2024. Photomaker: Customizing realistic human photos via stacked id embedding. In CVPR. 8640–8650

  17. [17]

    Zhuoying Li, Zhu Xu, Yuxin Peng, and Yang Liu. 2025. Balancing Preservation and Modification: A Region and Semantic Aware Metric for Instruction-Based Image Editing. arXiv:2506.13827 [cs.CV] https://arxiv.org/abs/2506.13827

  18. [18]

    Zhen Li, Zuo-Liang Zhu, Ling-Hao Han, Qibin Hou, Chun-Le Guo, and Ming- Ming Cheng. 2023. Amt: All-pairs multi-field transforms for efficient frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 9801–9810

  19. [19]

    Feng Liang, Haoyu Ma, Zecheng He, Tingbo Hou, Ji Hou, Kunpeng Li, Xiaoliang Dai, Felix Juefei-Xu, Samaneh Azadi, Animesh Sinha, et al. 2025. Movie Weaver: Tuning-Free Multi-Concept Video Personalization with Anchored Prompts. In Proceedings of the Computer Vision and Pattern Recognition Conference . 13146– 13156

  20. [20]

    Lijie Liu, Tianxiang Ma, Bingchuan Li, Zhuowei Chen, Jiawei Liu, Gen Li, Siyu Zhou, Qian He, and Xinglong Wu. 2025. Phantom: Subject-consistent video generation via cross-modal alignment. arXiv:2502.11079 [cs.CV] https://arxiv. org/abs/2502.11079

  21. [21]

    Dezhao Luo, Shaogang Gong, Jiabo Huang, Hailin Jin, and Yang Liu. 2024. Generative video diffusion for unseen cross-domain video moment retrieval. arXiv:2401.13329 [cs.CV] https://arxiv.org/abs/2401.13329

  22. [22]

    Ze Ma, Daquan Zhou, Xue-She Wang, Chun-Hsiao Yeh, Xiuyu Li, Huanrui Yang, Zhen Dong, Kurt Keutzer, and Jiashi Feng. 2024. Magic-me: Identity-specific video customized diffusion. In European Conference on Computer Vision . Springer, 19–37

  23. [23]

    Chaojie Mao, Jingfeng Zhang, Yulin Pan, Zeyinzi Jiang, Zhen Han, Yu Liu, and Jingren Zhou. 2025. Ace++: Instruction-based image creation and editing via context-aware content filling. arXiv:2501.02487 [cs.CV] https://arxiv.org/abs/ 2501.02487

  24. [24]

    MiniMax. 2024. Hailuo s2v-01. https://www.minimaxi.com/en/news/s2v-01- release/

  25. [25]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. 2021. Learning transferable visual models from natural language supervision. In International conference on machine learning . PmLR, 8748–8763

  26. [26]

    Yujun Shi, Jun Hao Liew, Hanshu Yan, Vincent YF Tan, and Jiashi Feng. 2024. InstaDrag: Lightning Fast and Accurate Drag-based Image Editing Emerging from Videos. arXiv:2405.13722 [cs.CV] https://arxiv.org/abs/2405.13722

  27. [27]

    Zhenyu Tang, Junwu Zhang, Xinhua Cheng, Wangbo Yu, Chaoran Feng, Yatian Pang, Bin Lin, and Li Yuan. 2025. Cycle3d: High-quality and consistent image-to- 3d generation via generation-reconstruction cycle. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 39. 7320–7328

  28. [28]

    Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. 2024. Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution. arXiv:2409.12191 [cs.CV] https://...

  29. [29]

    Zhao Wang, Aoxue Li, Enze Xie, Lingting Zhu, Yong Guo, Qi Dou, and Zhenguo Li. 2024. Customvideo: Customizing text-to-video generation with multiple subjects. arXiv:2401.09962 [cs.CV] https://arxiv.org/abs/2401.09962

  30. [30]

    Yujie Wei, Shiwei Zhang, Zhiwu Qing, Hangjie Yuan, Zhiheng Liu, Yu Liu, Yingya Zhang, Jingren Zhou, and Hongming Shan. 2024. Dreamvideo: Composing your dream videos with customized subject and motion. In CVPR. 6537–6549

  31. [31]

    Jianzong Wu, Xiangtai Li, Yanhong Zeng, Jiangning Zhang, Qianyu Zhou, Yining Li, Yunhai Tong, and Kai Chen. 2024. Motionbooth: Motion-aware customized text-to-video generation. Advances in Neural Information Processing Systems 37 (2024), 34322–34348

  32. [32]

    Zhu Xu, Qingchao Chen, Yuxin Peng, and Yang Liu. 2024. Semantic-aware human object interaction image generation. In Forty-first International Conference on Machine Learning

  33. [33]

    Zhu Xu, Zhaowen Wang, Yuxin Peng, and Yang Liu. 2025. Customized Human Object Interaction Image Generation. In Proceedings of the ACM International Conference on Multimedia (ACM-MM)

  34. [34]

    Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al

  35. [35]

    Wangbo Yu, Chaoran Feng, Jiye Tang, Xu Jia, Li Yuan, and Yonghong Tian. 2024. EvaGaussians: Event Stream Assisted Gaussian Splatting from Blurry Images. arXiv:2405.20224 [cs.CV] https://arxiv.org/pdf/2405.20224

  36. [36]

    Wangbo Yu, Jinbo Xing, Li Yuan, Wenbo Hu, Xiaoyu Li, Zhipeng Huang, Xi- angjun Gao, Tien-Tsin Wong, Ying Shan, and Yonghong Tian. 2024. ViewCrafter: Taming Video Diffusion Models for High-fidelity Novel View Synthesis. arXiv:2409.02048 [cs.CV] https://arxiv.org/abs/2409.02048

  37. [37]

    Shenghai Yuan, Xianyi He, Yufan Deng, Yang Ye, Jinfa Huang, Bin Lin, Chongyang Ma, Jiebo Luo, and Li Yuan. 2025. OpenS2V-Nexus: A Detailed Benchmark and Million-Scale Dataset for Subject-to-Video Generation. arXiv:2505.20292 [cs.CV] https://arxiv.org/abs/2505.20292

  38. [38]

    Shenghai Yuan, Jinfa Huang, Xianyi He, Yunyang Ge, Yujun Shi, Liuhan Chen, Jiebo Luo, and Li Yuan. 2025. Identity-preserving text-to-video generation by frequency decomposition. In Proceedings of the Computer Vision and Pattern Recognition Conference. 12978–12988

  39. [39]

    Shenghai Yuan, Jinfa Huang, Yujun Shi, Yongqi Xu, Ruijie Zhu, Bin Lin, Xinhua Cheng, Li Yuan, and Jiebo Luo. 2024. MagicTime: Time-lapse Video Generation Models as Metamorphic Simulators. arXiv:2404.05014 [cs.CV] https://arxiv.org/ abs/2404.05014

  40. [40]

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. 2023. Adding conditional control to text-to-image diffusion models. In ICCV. 3836–3847

  41. [41]

    Xin Zhang, Yanzhao Zhang, Wen Xie, Mingxin Li, Ziqi Dai, Dingkun Long, Pengjun Xie, Meishan Zhang, Wenjie Li, and Min Zhang. 2024. GME: Improving Universal Multimodal Retrieval by Multimodal LLMs. arXiv:2412.16855 [cs.CV] https://arxiv.org/abs/2412.16855

  42. [2024]

    arXiv:2408.06072 [cs.CV] https://arxiv.org/abs/2408.06072

    Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv:2408.06072 [cs.CV] https://arxiv.org/abs/2408.06072

  43. [2025]

    arXiv:2503.07598 [cs.CV] https://arxiv.org/abs/2503.07598

    Vace: All-in-one video creation and editing. arXiv:2503.07598 [cs.CV] https://arxiv.org/abs/2503.07598

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.