Pith. sign in

REVIEW 5 major objections 7 minor 20 references

Vera claims that identity-faithful human video generation is achievable by pairing cross-clip identity-aligned training data with face-focused supervision and layer-aware reference attention.

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 →

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

T0 review · deepseek-v4-flash

2026-08-01 10:20 UTC pith:IO2NVDWJ

load-bearing objection Vera is a plausible engineering contribution with a nice cross-clip data idea, but the evaluation doesn't yet rule out memorization or selection, so treat the numbers as provisional. the 5 major comments →

arxiv 2607.20247 v1 pith:IO2NVDWJ submitted 2026-07-22 cs.CV

Vera: Identity-Faithful Human Subject-to-Video Generation

Yulong Xu , Xinyue Liu , Shujuan Li , huafeng shi , Yan Zhou , Jiwen Liu , Xintao Wang , Yu Shen Liu
show 1 more author
Huaibo Huang
This is my paper
classification cs.CV
keywords subject-to-video generationidentity preservationmulti-person video generationdiffusion transformercross-clip retrievalface-focused masked supervisionreference-aware attentionvideo personalization
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 paper argues that human identity in AI-generated video fails because training data lacks per-person identity correspondence and because standard losses and attention treat faces like any other image region. Vera addresses this with a million-pair dataset in which references are retrieved from different clips of the same person, then adds two training-time mechanisms: masked supervision that up-weights face regions and layer-wise attention control that keeps reference tokens clean while boosting video-to-reference identity readout at middle layers. If the results hold, Vera gives a measurable improvement in facial identity consistency and multi-person role binding over generic subject-to-video baselines.

Core claim

Vera claims that identity-preserving human video generation can be achieved without changing inference by (1) training on identity-aligned cross-clip image-video pairs, so the model must learn identity rather than copy the same frames; (2) applying Identity-Focal Masked Supervision, which weights face regions roughly twice as much as background in the rectified-flow training loss and masks subtitle artifacts; and (3) applying Reference-Aware Layer-wise Attention, which blocks reference-to-video attention in shallow layers of the transformer backbone and adds an identity residual in middle layers. On a curated 100-identity benchmark spanning single- and multi-person clips, Vera reports FaceSi

What carries the argument

The load-bearing object is the identity-aligned dataset: each training sample pairs a target video clip with reference images of the same person retrieved from other clips, filtered to a moderate similarity band so references are identity-consistent but not near-duplicates. On top of this, IFMS is a spatially weighted training loss that concentrates supervision on face regions, and RALA is a two-part attention control — a mask that protects reference tokens from video noise in shallow layers and a residual that amplifies video-to-reference identity readout at identity-sensitive middle layers.

Load-bearing premise

The data pipeline assumes that automated face clustering and cross-clip similarity search reliably match the same person across different clips; the paper validates this only by visual inspection, not by measuring clustering or retrieval accuracy.

What would settle it

Build a labeled set of multi-person videos with known identities, run the dataset pipeline end-to-end, and measure identity-retrieval precision. If a substantial fraction of retrieved references belong to a different person, then the model is being trained to preserve the wrong face and the claimed identity-fidelity gains are not attributable to the true identity signal.

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

If this is right

  • Training on cross-clip references prevents the model from copying target-clip pose and background, pushing it to learn genuine identity correspondence.
  • IFMS provides stronger supervision on identity-critical regions with zero inference overhead, so it can be added to existing training pipelines.
  • RALA stabilizes reference conditioning while leaving motion generation flexible, reducing both identity drift and excessive copying from reference images.
  • A single unified framework handles both single-person and multi-person generation, improving identity-role binding and reducing subject confusion and attribute swapping.
  • The reported gains hold alongside naturalness and prompt alignment, not at their expense.

Where Pith is reading between the lines

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

  • If cross-clip retrieval precision is lower than assumed, part of the measured identity gain may be the model learning a distorted 'average' face rather than the true identity; a labeled retrieval audit would settle this.
  • The same data-and-supervision recipe could transfer to other identity-critical domains such as animals or avatars, though the similarity band thresholds would need re-estimation for each class.
  • RALA's layer-dependent schedule suggests a broader design principle: reference conditioning in generative transformers should be asymmetric across depth, which could apply beyond human video to any reference-conditioned generation task.
  • The 100-identity benchmark may under-sample visually similar faces; stress-testing on lookalike pairs would show whether the identity gains persist under maximal confusion pressure.

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

5 major / 7 minor

Summary. The paper proposes Vera, a human-centric subject-to-video (S2V) generation framework that combines two main contributions: (1) a data pipeline that constructs a large corpus of identity-aligned image-video pairs via face clustering and cross-clip retrieval, and (2) two training-time components -- Identity-Focal Masked Supervision (IFMS), which reweights the flow-matching loss toward facial regions, and Reference-Aware Layer-wise Attention (RALA), which modifies reference-video attention in the DiT backbone to protect reference tokens and reinforce identity readout in certain layers. The method is fine-tuned from Wan-2.2-TI2V-5B and evaluated on a self-curated benchmark of 100 identities with 600 generated videos, comparing against VACE, BindWeave, and Phantom. The paper reports improved face-similarity metrics (FaceSim-Arc 0.571 vs. 0.531 for the strongest baseline, FaceSim-Cur 0.529 vs. 0.507) while maintaining competitive motion and prompt-following scores, and includes qualitative results for single- and multi-person examples.

Significance. If the claims hold, the contributions are of interest to the video generation community: the cross-clip data pipeline directly addresses the shortage of identity-aligned S2V training data, and both IFMS and RALA are simple, training-side modifications that require no inference overhead. The internal ablations are consistent with the design intentions, and the paper is transparent about the base model and hyperparameter settings. However, the empirical evidence is currently insufficient to support the central generalization claim: the benchmark is self-curated with no stated identity-disjointness from the training set, no error bars or statistical tests are provided, identity-specific baselines are absent, and the data pipeline itself is not quantitatively validated. Thus the work is a solid proof-of-concept whose significance is currently limited by these evaluation gaps.

major comments (5)
  1. [Sec. 3.1 / Appendix C.2 / F] The benchmark of 100 identities is not stated to be disjoint from the approximately 1M training pairs produced by the pipeline. Since the benchmark is self-curated and unreleased, and the training data is built from raw videos with face embeddings, benchmark identities may overlap with the training set, inflating the FaceSim scores in Table 1 (0.571 vs. 0.531) via memorization rather than identity preservation for unseen people. This is load-bearing because the practical S2V claim is zero-shot personalization. Please provide an identity-overlap audit (e.g., maximum face similarity between each benchmark reference and the training set identities), and either exclude overlapping identities or report results on a verified held-out subset.
  2. [Sec. 4.1 / Tables 1-4] All quantitative results are single numbers with no error bars or significance tests, and hyperparameters (e.g., wface in Table 3, wmid in Table 4) are tuned on the same benchmark. The reported gains are small (e.g., FaceSim-Arc 0.571 vs. 0.531; FaceSim-Cur 0.529 vs. 0.507), and could lie within run-to-run noise across the 600 generated videos. Please report per-seed results, standard deviations, and/or statistical significance tests, and ideally select hyperparameters on a separate validation set before evaluating the final model on held-out test cases.
  3. [Sec. 4.2 / Related Work] The paper cites identity-specific human video generation methods (ConsisID, ID-Animator) but does not compare against them in Table 1, using only generic S2V baselines (VACE, BindWeave, Phantom). Since the paper's central claim is about facial identity consistency rather than generic subject preservation, the evaluation should include at least one identity-specific baseline, or provide a clear justification for their exclusion. Without this, the claim that Vera 'improves human identity consistency' is only established relative to generic S2V methods.
  4. [Appendix B.3, Eq. (20)] The data pipeline depends on intra-clip identity clustering and cross-clip retrieval using face embeddings, but the only validation described is manual 'visual verification.' If the clustering or retrieval associates the wrong identity with a target clip, the model is trained to preserve the wrong face, directly undermining the central identity-fidelity claim. Please provide quantitative validation of the pipeline, such as cluster purity or retrieval precision@k on a labeled subset, and report representative failure cases.
  5. [Sec. 3.3 / Appendix C.1] The RALA specification is incomplete. Eq. (5) defines a residual coefficient β_l, but Appendix C.1 instead describes a multiplier γ_l=1.5 and states this is 'equivalent to' β_l=0.5, which is notationally inconsistent and not immediately derivable from the formula. The exact layer indices for the shallow layers L_s and the middle-layer band are not reported for the 5B DiT, making the method hard to reproduce. Additionally, the paper claims 'our experimental analysis further shows that video-to-reference identity readout exhibits a clear layer-dependent pattern,' but no such analysis is shown. Please add the layer-wise analysis, specify the exact layer schedule, and correct the notation.
minor comments (7)
  1. [Sec. 4.2] The text says 'as reported in Tab. 3' but the quantitative comparison is Table 1. Please correct the cross-reference.
  2. [Appendix B.3, Eq. (21)] The diversity term sums over Ra,Rb in S without an explicit normalizer; using |S|(|S|-1) or a similar term would make the objective scale-invariant with respect to the number of selected references.
  3. [Appendix B.3] The threshold values τ_low and τ_high in Eq. (20) are not reported anywhere. Please provide the concrete values or at least the similarity ranges used in the final dataset.
  4. [Sec. 3.2 / Fig. 5] The paper repeatedly claims that IFMS 'significantly reduces subtitle-induced artifacts,' but no quantitative metric or dedicated qualitative comparison for subtitle suppression is provided. Please add evidence or soften the claim.
  5. [Appendix C.1] The terms 'face scale=1.0' and 'face offset=1.0' are undefined; please explain how the face bounding boxes are enlarged and rasterized.
  6. [Table 1] VACE-1.3B has the highest MotionSmoothness but the text only says 'reasonable temporal smoothness' for Vera. Please discuss this trade-off explicitly.
  7. [References] The 'OpenS2V-Nexus' work is cited twice (Yuan et al. 2025a and Yuan et al. 2026); the duplicate should be merged or disambiguated.

Circularity Check

0 steps flagged

No significant circularity: the training objective, data curation, and evaluation are not equivalent by construction.

full rationale

Vera is an empirical method, not a formal derivation whose output reduces to a fitted input. The identity-aligned dataset is constructed by clustering face embeddings and retrieving cross-clip references (Eqs. 19-20), which is a data-construction choice rather than a prediction. IFMS is a spatially weighted denoising loss (Eqs. 1-2) and RALA is an attention mask plus residual readout (Eqs. 3-5); neither is defined in terms of the evaluation metrics, and the paper's reported FaceSim gains are not mathematical consequences of the loss definitions. The ablations tune hyperparameters on the same self-curated benchmark, but that is standard model selection, not a fitted parameter being renamed as a prediction. The benchmark is self-created, and possible identity overlap with the training set is a legitimate external-validity/correctness concern, but the paper provides no equation or construction that forces the reported identity gains. Self-citations, including the Wan base model and ConceptMaster, are used as infrastructure or related work, not as load-bearing evidence for the identity-preservation claim. Appendix G honestly lists limitations without revealing any circular step. Overall, no claim reduces to its own inputs by construction.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

The method introduces no new physical entities. Its free parameters are primarily loss weights and data-pipeline thresholds, several of which are tuned directly against the evaluation metrics. The most expensive assumptions are about the reliability of face-identity clustering and retrieval, and about the layer-localized behavior of identity attention.

free parameters (6)
  • wface / wbg loss weights = wface=2.0, wbg=1.0
    Selected by ablation on the evaluation benchmark (Table 3); wface=2.0 maximizes FaceSim and NaturalScore.
  • wmid / beta_l identity residual weight = wmid=1.5, beta_l=0.5
    Selected by ablation on the evaluation benchmark (Table 4); larger values degrade motion smoothness.
  • Cross-clip similarity thresholds tau_low, tau_high = not reported
    Data pipeline filters candidate references by a moderate similarity band (Eq. 20); values are chosen by hand and not specified.
  • Reference diversity weight lambda = not reported
    Balances identity reliability against visual diversity in reference selection (Eq. 21); value not given.
  • Number of shallow RAPM layers Ls = not reported
    Controls how many early DiT layers block reference-to-video attention (Eq. 3); no value or selection procedure is provided.
  • Face bounding-box enlargement and face scale/offset = 50% enlargement; face scale=1.0, face offset=1.0
    Used to rasterize face masks in latent space (Appendix C.1); chosen without reported sensitivity analysis.
axioms (5)
  • standard math Rectified-flow training objective is a valid base for fine-tuning video DiT models.
    Invoked in Appendix A Eq. 6-9 without proof; standard in the field.
  • domain assumption ArcFace and CurricularFace embeddings measure identity fidelity in generated videos.
    FaceSim metrics treat face-recognition embedding similarity as ground truth for identity preservation (Section 4.1).
  • domain assumption Face detection, embedding, and clustering yield correct person identities across clips.
    The entire dataset construction depends on this; the paper provides only manual visual verification (Appendix B).
  • ad hoc to paper Identity readout is concentrated in middle layers of the DiT backbone.
    RALA assumes a layer-dependent readout pattern based on 'our experimental analysis' (Section 3.3), but the analysis is not shown.
  • domain assumption Wan-2.2-TI2V-5B is an adequate base model for identity-focused fine-tuning.
    All experiments fine-tune one base model; no comparison with training from another backbone is provided.

pith-pipeline@v1.3.0-alltime-deepseek · 17206 in / 11091 out tokens · 94086 ms · 2026-08-01T10:20:49.121642+00:00 · methodology

0 comments
read the original abstract

Subject-to-video (S2V) generation has made substantial progress in preserving reference subjects across diverse categories, yet generic subject consistency remains insufficient for human-centric generation. A video may appear globally consistent while identity-critical human details still drift across frames, poses, and interactions. This issue becomes more severe in multi-person scenarios, where incorrect identity-role binding leads to subject confusion, attribute swapping, and excessive copying of reference-specific appearance cues. We propose Vera, a unified human-centric S2V framework for single- and multi-person generation. We first construct a million-pair identity-aligned human image-video dataset through person-level cross-clip retrieval, providing explicit identity correspondence and diverse references. Built on this dataset, Vera introduces two complementary designs. Identity-Focal Masked Supervision (IFMS) strengthens identity-aware learning with spatially focused supervision while reducing interference from irrelevant artifacts. Reference-Aware Layer-wise Attention (RALA) regulates how video tokens interact with reference identity cues in the DiT backbone, preserving stable identity anchors and enhancing layer-aware identity readout. Extensive experiments demonstrate that Vera improves human identity consistency, multi-person subject binding, and motion naturalness, while reducing identity confusion and excessive reference-image copying.

Figures

Figures reproduced from arXiv: 2607.20247 by Huafeng Shi, Huaibo Huang, Jiwen Liu, Shujuan Li, Xintao Wang, Xinyue Liu, Yan Zhou, Yulong Xu, Yu Shen Liu.

Figure 1
Figure 1. Figure 1: Data processing pipeline for identity-consistent video generation. The process includes [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the Vera framework. Text prompts, reference image sets, and input videos are [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison on multi-person subject-to-video generation. Our method better [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative ablation of IFMS and RALA. IFMS improves overall facial consistency, while [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: More qualitative results of Vera. Each case contains the input reference image(s), generated [PITH_FULL_IMAGE:figures/full_fig_p020_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Statistics of the identity-aligned human image–video dataset. The dataset contains 1,001,891 [PITH_FULL_IMAGE:figures/full_fig_p021_7.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

20 extracted references · 16 linked inside Pith

  1. [1]

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

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

    Pith/arXiv arXiv

  2. [4]

    An image is worth one word: Personalizing text-to-image generation using textual inversion.arXiv preprint arXiv:2208.01618,

    Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion.arXiv preprint arXiv:2208.01618,

    Pith/arXiv arXiv

  3. [9]

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

    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,

    Pith/arXiv arXiv

  4. [10]

    Personalvideo: High id-fidelity video customization without dynamic and semantic degradation

    Hengjia Li, Haonan Qiu, Shiwei Zhang, Xiang Wang, Yujie Wei, Zekun Li, Yingya Zhang, Boxi Wu, and Deng Cai. Personalvideo: High id-fidelity video customization without dynamic and semantic degradation. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 19406–19416, 2025a. Zhaoyang Li, Dongjun Qian, Kai Su, Qishuai Diao, Xiang...

    arXiv

  5. [12]

    Make-a-video: Text-to-video generation without text-video data.arXiv preprint arXiv:2209.14792,

    Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data.arXiv preprint arXiv:2209.14792,

    Pith/arXiv arXiv

  6. [13]

    Wan: Open and advanced large-scale video generative models

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

    Pith/arXiv arXiv

  7. [14]

    Instantid: Zero-shot identity-preserving generation in seconds.arXiv preprint arXiv:2401.07519,

    Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, Anthony Chen, Huaxia Li, Xu Tang, and Yao Hu. Instantid: Zero-shot identity-preserving generation in seconds.arXiv preprint arXiv:2401.07519,

    Pith/arXiv arXiv

  8. [15]

    Stand-in: A lightweight and plug-and-play identity control for video generation.arXiv preprint arXiv:2508.07901,

    11 Bowen Xue, Zheng-Peng Duan, Qixin Yan, Wenjing Wang, Hao Liu, Chun-Le Guo, Chongyi Li, Chen Li, and Jing Lyu. Stand-in: A lightweight and plug-and-play identity control for video generation.arXiv preprint arXiv:2508.07901,

    arXiv

  9. [16]

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

    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,

    Pith/arXiv arXiv

  10. [17]

    Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models.arXiv preprint arXiv:2308.06721,

    Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models.arXiv preprint arXiv:2308.06721,

    Pith/arXiv arXiv

  11. [18]

    Opens2v-nexus: A detailed benchmark and million-scale dataset for subject-to-video generation

    Shenghai Yuan, Xianyi He, Yufan Deng, Yang Ye, Jinfa Huang, Bin Lin, Jiebo Luo, and Li Yuan. Opens2v-nexus: A detailed benchmark and million-scale dataset for subject-to-video generation. arXiv preprint arXiv:2505.20292, 2025a. Shenghai Yuan, Jinfa Huang, Xianyi He, Yunyang Ge, Yujun Shi, Liuhan Chen, Jiebo Luo, and Li Yuan. Identity-preserving text-to-vi...

    Pith/arXiv arXiv

  12. [19]

    Open-sora: Democratizing efficient video production for all.arXiv preprint arXiv:2412.20404,

    Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all.arXiv preprint arXiv:2412.20404,

    Pith/arXiv arXiv

  13. [20]

    13 B Data Construction Pipeline

    12 Appendix A Preliminaries. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13 B Data Construction Pipeline. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14 B.1 Video Segmentation and Face-Level Fi...

    arXiv

  14. [2000]

    ∆-DiT: A training-free acceleration method tailored for diffusion transformers.arXiv preprint arXiv:2406.01125,

    Pengtao Chen, Mingzhu Shen, Peng Ye, Jianjian Cao, Chongjun Tu, Christos-Savvas Bouganis, Yiren Zhao, and Tao Chen. ∆-DiT: A training-free acceleration method tailored for diffusion transformers.arXiv preprint arXiv:2406.01125,

    Pith/arXiv arXiv

  15. [2020]

    Conceptmaster: Multi-concept video customization on diffusion transformer models without test-time tuning.arXiv preprint arXiv:2501.04698,

    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,

    Pith/arXiv arXiv

  16. [2022]

    Wildactor: Unconstrained identity-preserving video generation.arXiv preprint arXiv:2603.00586,

    Qin Guo, Tianyu Yang, Xuanhua He, Fei Shen, Yong Zhang, Zhuoliang Kang, Xiaoming Wei, and Dan Xu. Wildactor: Unconstrained identity-preserving video generation.arXiv preprint arXiv:2603.00586,

    arXiv

  17. [2023]

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

    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,

    Pith/arXiv arXiv

  18. [2024]

    Id- animator: Zero-shot identity-preserving human video generation.arXiv preprint arXiv:2404.15275,

    Xuanhua He, Quande Liu, Shengju Qian, Xin Wang, Tao Hu, Ke Cao, Keyu Yan, and Jie Zhang. Id- animator: Zero-shot identity-preserving human video generation.arXiv preprint arXiv:2404.15275,

    Pith/arXiv arXiv

  19. [2025]

    Skyreels-a2: Compose anything in video diffusion transformers

    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,

    Pith/arXiv arXiv

  20. [2026]

    Animatediff: Animate your personalized text-to-image diffusion models without specific tuning.arXiv preprint arXiv:2307.04725,

    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.arXiv preprint arXiv:2307.04725,

    Pith/arXiv arXiv