pith. sign in

arxiv: 2606.02441 · v1 · pith:LHHIIPK4new · submitted 2026-06-01 · 💻 cs.CV

Spatial-Temporal Decoupled Reference Conditioning for Identity-Preserving Text-to-Video Generation

Yuheng Chen , Teng Hu , Yuji Wang , Qingdong He , Lizhuang Ma , Jiangning Zhang This is my paper

Pith reviewed 2026-06-28 15:23 UTC · model grok-4.3

classification 💻 cs.CV
keywords identity-preserving video generationreference conditioningtext-to-video generationRoPE positional encodinglatent feature injectionclassifier-free guidancediffusion modelstemporal consistency
0
0 comments X

The pith

ST-DRC decouples spatial and temporal reference signals so identity details reach video generation through attention rather than pixel copying.

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

The paper proposes a framework that encodes a reference image with the video VAE and concatenates it to noisy video latents, giving the model direct access to low-level identity features. It then applies a modified positional encoding called TASS-RoPE that keeps reference tokens temporally close to the video but shifts them spatially, so identity information travels through spatio-temporal attention layers instead of enabling direct copy-paste. Appearance-invariant augmentations and face-guided losses further prevent the diffusion objective from diluting identity supervision, while a three-stream guidance method at inference separately tunes text adherence and reference fidelity.

Core claim

By performing latent in-context feature injection and introducing TASS-RoPE to place reference tokens near the video sequence in time but shifted in space, the model allows reference information to flow through spatio-temporal attention while suppressing pixel-level copy-paste shortcuts; combining this with appearance-invariant reference augmentation and face-guided identity objectives strengthens identity preservation under the diffusion training process.

What carries the argument

TASS-RoPE, a Temporal-Adjacent Spatial-Shifted RoPE scheme that positions reference tokens temporally adjacent yet spatially offset to route identity signals through attention.

If this is right

  • Identity preservation improves without requiring extra adapter modules.
  • Prompt alignment and temporal consistency remain high alongside identity fidelity.
  • Three-stream classifier-free guidance allows independent control of text and reference strength at inference.
  • The approach works as a lightweight addition to an existing video diffusion backbone.

Where Pith is reading between the lines

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

  • The same latent concatenation plus shifted positional encoding could be tested on non-human reference subjects such as objects or scenes.
  • The decoupling might reduce the need for heavy identity-specific fine-tuning in other generative video tasks.
  • Extending the spatial shift distance could be explored to handle longer reference-to-video time gaps.

Load-bearing premise

Spatial shifting of reference tokens in the positional encoding prevents direct pixel copying while still permitting useful identity information to propagate through attention.

What would settle it

Generating videos after removing the spatial shift component of TASS-RoPE and measuring whether copy-paste artifacts increase or identity fidelity drops.

Figures

Figures reproduced from arXiv: 2606.02441 by Jiangning Zhang, Lizhuang Ma, Qingdong He, Teng Hu, Yuheng Chen, Yuji Wang.

Figure 1
Figure 1. Figure 1: Examples of identity-preserving text-to-video generation by our ST-DRC. Given a reference face, our method generates [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of ST-DRC. (a) The reference image is encoded into the video latent space and concatenated with noisy [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison on VIP-200K [67]. Given the same reference image and text prompt, ST-DRC preserves the target identity more faithfully while producing prompt-consistent and visually coherent videos compared with baselines. where 𝑚𝑓 indicates whether a valid face is detected in frame 𝑓 . To reduce cross-frame identity drift, we further add a temporal identity consistency loss: 𝑒¯ = Í 𝑓 ∈ F 𝑚𝑓 𝑒𝑓 Í 𝑓 … view at source ↗
read the original abstract

Identity-preserving video generation (IPVG) aims to synthesize high-fidelity videos that follow text prompts while faithfully preserving a reference identity. Despite recent progress, existing IPVG methods still struggle to balance high-level semantic control and low-level identity fidelity. To bridge this gap, we propose ST-DRC, an effective Spatial-Temporal Decoupled Reference Conditioning framework for identity-preserving text-to-video generation. At the framework level, ST-DRC performs latent in-context feature injection by encoding the reference image with the video VAE and concatenating it with noisy video latents, enabling rich low-level identity details to be accessed without additional adapters. To separate identity-aware reference retrieval from appearance copying, we introduce TASS-RoPE, a Temporal-Adjacent Spatial-Shifted RoPE scheme that places reference tokens near the video sequence in time but shifts them in space, allowing reference information to flow through spatio-temporal attention while suppressing pixel-level copy-paste shortcuts. To further prevent shortcut learning and strengthen the otherwise diluted identity supervision in the diffusion objective, we combine appearance-invariant reference augmentation with face-guided identity objectives, encouraging the model to preserve identity under variations in color, pose, and layout. At inference time, we introduce a three-stream reference classifier-free guidance strategy that independently controls text adherence and reference fidelity. Experiments demonstrate that ST-DRC achieves strong identity preservation, prompt alignment, temporal consistency, and video quality with a lightweight design built on LTX-2.3. Our method ranks among the top submissions in the facial identity-preserving video generation track, validating the effectiveness of spatial-temporal decoupled reference conditioning.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 0 minor

Summary. The paper introduces ST-DRC, a Spatial-Temporal Decoupled Reference Conditioning framework for identity-preserving text-to-video generation. It encodes a reference image via the video VAE and concatenates it with noisy video latents for in-context injection; introduces TASS-RoPE (Temporal-Adjacent Spatial-Shifted RoPE) to place reference tokens temporally adjacent but spatially shifted so that spatio-temporal attention carries high-level identity semantics while blocking low-level copy-paste; adds appearance-invariant augmentations and face-guided identity losses; and uses three-stream classifier-free guidance at inference. The method is built on LTX-2.3 and is reported to achieve strong identity preservation, prompt alignment, temporal consistency and video quality, ranking among the top entries in a facial identity-preserving video generation track.

Significance. If the experimental claims hold and the mechanistic assumptions of TASS-RoPE are validated, the work would supply a lightweight, adapter-free approach to balancing semantic control and low-level fidelity in conditional video diffusion. The spatial-temporal decoupling idea could transfer to other reference-conditioned generation settings. The manuscript does not supply machine-checked proofs, parameter-free derivations or open reproducible code, so its primary contribution remains empirical.

major comments (2)
  1. [Abstract, §3] Abstract and §3: the central mechanistic claim that TASS-RoPE 'suppresses pixel-level copy-paste shortcuts' by spatial shifting while still permitting semantic flow is asserted without supporting evidence. No attention-map visualizations, no isolated ablation that removes only the spatial-shift component, and no analysis of VAE latent correlation distances are provided to show that the chosen shift distance disrupts exact token alignment more than semantic retrieval.
  2. [Abstract, §4–5] Abstract (and presumably §4–5): the claim that 'ST-DRC achieves strong identity preservation, prompt alignment, temporal consistency, and video quality' and 'ranks among the top submissions' is stated without any quantitative metrics, baselines, ablation tables, or error analysis. This prevents evaluation of the magnitude of improvement or the contribution of each proposed component.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We address the two major comments below and will revise the manuscript accordingly to strengthen the mechanistic evidence and experimental reporting.

read point-by-point responses

  1. Referee: [Abstract, §3] Abstract and §3: the central mechanistic claim that TASS-RoPE 'suppresses pixel-level copy-paste shortcuts' by spatial shifting while still permitting semantic flow is asserted without supporting evidence. No attention-map visualizations, no isolated ablation that removes only the spatial-shift component, and no analysis of VAE latent correlation distances are provided to show that the chosen shift distance disrupts exact token alignment more than semantic retrieval.

    Authors: We agree that direct supporting evidence for the TASS-RoPE mechanism is currently missing from the manuscript. In the revision we will add (i) attention-map visualizations comparing reference-to-video attention with and without the spatial shift, (ii) an isolated ablation that removes only the spatial-shift component while keeping temporal adjacency, and (iii) quantitative analysis of VAE latent correlation distances at varying shift offsets. These additions will be placed in a new subsection of §3. revision: yes

  2. Referee: [Abstract, §4–5] Abstract (and presumably §4–5): the claim that 'ST-DRC achieves strong identity preservation, prompt alignment, temporal consistency, and video quality' and 'ranks among the top submissions' is stated without any quantitative metrics, baselines, ablation tables, or error analysis. This prevents evaluation of the magnitude of improvement or the contribution of each proposed component.

    Authors: We acknowledge that the current manuscript version relies primarily on qualitative results and the competition ranking without accompanying quantitative tables. In the revised version we will expand §4 and §5 with (i) numerical metrics for identity preservation (e.g., ArcFace cosine similarity), prompt alignment (CLIP text-video scores), temporal consistency (e.g., frame-wise optical flow consistency), and perceptual quality, (ii) comparisons against relevant baselines, (iii) component-wise ablation tables, and (iv) error analysis across prompt categories. These tables will directly support the claims made in the abstract. revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical architectural proposal

full rationale

The paper proposes ST-DRC and TASS-RoPE as an architectural framework for identity-preserving video generation, validated through experiments on LTX-2.3. No mathematical derivations, equations, parameter fittings, or self-referential definitions are present in the abstract or described claims. The central mechanism (spatial shift in RoPE) is an ansatz justified by design intent and empirical results rather than reducing to prior inputs by construction. No load-bearing self-citations or uniqueness theorems are invoked. This matches the default expectation for non-circular empirical work.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The approach rests on standard diffusion-model assumptions about latent-space conditioning and attention mechanisms; no new free parameters, axioms, or invented entities are introduced or quantified in the abstract.

axioms (1)
  • domain assumption Encoding the reference image with the video VAE and concatenating it with noisy video latents enables rich low-level identity details to be accessed without additional adapters.
    Directly stated as the mechanism for low-level identity access.

pith-pipeline@v0.9.1-grok · 5834 in / 1266 out tokens · 33507 ms · 2026-06-28T15:23:44.110652+00:00 · methodology

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

70 extracted references · 31 canonical work pages · 12 internal anchors

  1. [1]

    Hila Chefer, Uriel Singer, Amit Zohar, Yuval Kirstain, Adam Polyak, Yaniv Taig- man, Lior Wolf, and Shelly Sheynin. 2025. Videojam: Joint appearance-motion representations for enhanced motion generation in video models.arXiv preprint arXiv:2502.02492(2025)

    work page arXiv 2025

  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

    2024

  3. [3]

    Guibin Chen, Dixuan Lin, Jiangping Yang, Chunze Lin, Junchen Zhu, Mingyuan Fan, Hao Zhang, Sheng Chen, Zheng Chen, Chengcheng Ma, et al. 2025. Skyreels- v2: Infinite-length film generative model.arXiv preprint arXiv:2504.13074(2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  4. [4]

    Hong Chen, Xin Wang, Guanning Zeng, Yipeng Zhang, Yuwei Zhou, Feilin Han, Yaofei Wu, and Wenwu Zhu. 2025. Videodreamer: Customized multi-subject text- to-video generation with disen-mix finetuning on language-video foundation models.IEEE Transactions on Multimedia(2025)

    2025

  5. [5]

    Jingxi Chen, Zongxia Li, Zhichao Liu, Guangyao Shi, Xiyang Wu, Fuxiao Liu, Cornelia Fermuller, Brandon Y Feng, and Yiannis Aloimonos. 2025. First Frame Is the Place to Go for Video Content Customization.arXiv preprint arXiv:2511.15700 (2025)

    work page arXiv 2025

  6. [6]

    Liyang Chen, Tianxiang Ma, Jiawei Liu, Bingchuan Li, Zhuowei Chen, Lijie Liu, Xu He, Gen Li, Qian He, and Zhiyong Wu. 2025. Humo: Human-centric video generation via collaborative multi-modal conditioning.arXiv preprint arXiv:2509.08519(2025)

    work page arXiv 2025

  7. [7]

    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

    2025

  8. [8]

    Yuheng Chen, Qingdong He, Teng Hu, Yuji Wang, Yabiao Wang, Lizhuang Ma, and Jiangning Zhang. 2026. Omni-Customizer: End-to-End MultiModal Cus- tomization for Joint Audio-Video Generation.arXiv preprint arXiv:2605.17488 (2026)

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  9. [9]

    Zhuowei Chen, Bingchuan Li, Tianxiang Ma, Lijie Liu, Mingcong Liu, Yi Zhang, Gen Li, Xinghui Li, Siyu Zhou, Qian He, et al. 2025. Phantom-data: Towards a gen- eral subject-consistent video generation dataset.arXiv preprint arXiv:2506.18851 (2025)

    work page arXiv 2025

  10. [10]

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

    2019

  11. [11]

    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. InForty- first international conference on machine learning

    2024

  12. [12]

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

    work page arXiv 2025

  13. [13]

    Jiayi Gao, Changcheng Hua, Qingchao Chen, Yuxin Peng, and Yang Liu. 2025. Identity-Preserving Text-to-Video Generation via Training-Free Prompt, Image, and Guidance Enhancement. InProceedings of the 33rd ACM International Confer- ence on Multimedia. 13751–13757

    2025

  14. [14]

    Yoav HaCohen, Benny Brazowski, Nisan Chiprut, Yaki Bitterman, Andrew Kvochko, Avishai Berkowitz, Daniel Shalem, Daphna Lifschitz, Dudu Moshe, Eitan Porat, et al. 2026. LTX-2: Efficient Joint Audio-Visual Foundation Model. arXiv preprint arXiv:2601.03233(2026)

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  15. [15]

    Junjie He, Yifeng Geng, and Liefeng Bo. 2025. Uniportrait: A unified frame- work for identity-preserving single-and multi-human image personalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision. 14399– 14408

    2025

  16. [16]

    Xuanhua He, Quande Liu, Shengju Qian, Xin Wang, Tao Hu, Ke Cao, Keyu Yan, and Jie Zhang. 2024. Id-animator: Zero-shot identity-preserving human video generation.arXiv preprint arXiv:2404.15275(2024)

    work page arXiv 2024

  17. [17]

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

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  18. [18]

    Li Hu, Guangyuan Wang, Zhen Shen, Xin Gao, Dechao Meng, Lian Zhuo, Peng Zhang, Bang Zhang, and Liefeng Bo. 2025. Animate anyone 2: High-fidelity character image animation with environment affordance. InProceedings of the IEEE/CVF International Conference on Computer Vision. 10207–10217

    2025

  19. [19]

    Teng Hu, Zhentao Yu, Guozhen Zhang, Zihan Su, Zhengguang Zhou, You- liang Zhang, Yuan Zhou, Qinglin Lu, and Ran Yi. 2025. Harmony: Harmoniz- ing Audio and Video Generation through Cross-Task Synergy.arXiv preprint arXiv:2511.21579(2025)

    work page arXiv 2025

  20. [20]

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

    work page arXiv 2025

  21. [21]

    Teng Hu, Zhentao Yu, Zhengguang Zhou, Jiangning Zhang, Yuan Zhou, Qinglin Lu, and Ran Yi. 2026. PolyVivid: Vivid Multi-Subject Video Generation with Cross-Modal Interaction and Enhancement.Advances in Neural Information Processing Systems38 (2026), 49394–49420

    2026

  22. [22]

    Teng Hu, Jiangning Zhang, Zihan Su, and Ran Yi. 2026. UltraGen: High-Resolution Video Generation with Hierarchical Attention. InProceedings of the AAAI Con- ference on Artificial Intelligence, Vol. 40. 4923–4931

    2026

  23. [23]

    Chi-Pin Huang, Yen-Siang Wu, Hung-Kai Chung, Kai-Po Chang, Fu-En Yang, and Yu-Chiang Frank Wang. 2025. Videomage: Multi-subject and motion customiza- tion of text-to-video diffusion models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 17603–17612

    2025

  24. [24]

    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. Inproceedings of the IEEE/CVF conference on computer vision and pattern recognition. 5901–5910

    2020

  25. [25]

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

    work page arXiv 2025

  26. [26]

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

    2024

  27. [27]

    InsightFace Contributors. 2023. InsightFace: 2D and 3D Face Analysis Project. https://github.com/deepinsight/insightface

    2023

  28. [28]

    Yuming Jiang, Tianxing Wu, Shuai Yang, Chenyang Si, Dahua Lin, Yu Qiao, Chen Change Loy, and Ziwei Liu. 2024. Videobooth: Diffusion-based video generation with image prompts. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 6689–6700

    2024

  29. [29]

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

  30. [30]

    InProceedings of the IEEE/CVF International Conference on Computer Vision

    Vace: All-in-one video creation and editing. InProceedings of the IEEE/CVF International Conference on Computer Vision. 17191–17202

  31. [31]

    Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. 2021. Musiq: Multi-scale image quality transformer. InProceedings of the IEEE/CVF international conference on computer vision. 5148–5157

    2021

  32. [32]

    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 preprint arXiv:2412.03603 (2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  33. [33]

    LAION-AI. 2022. LAION-Aesthetics Predictor. https://github.com/LAION-AI/ aesthetic-predictor. A linear estimator on top of CLIP for predicting image aesthetic quality

    2022

  34. [34]

    Zhaoyang Li, Dongjun Qian, Kai Su, Qishuai Diao, Xiangyang Xia, Chang Liu, Wenfei Yang, Tianzhu Zhang, and Zehuan Yuan. 2025. Bindweave: Subject-consistent video generation via cross-modal integration.arXiv preprint arXiv:2510.00438(2025)

    work page arXiv 2025

  35. [35]

    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. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 9801–9810

    2023

  36. [36]

    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

    2025

  37. [37]

    Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le

  38. [38]

    Flow matching for generative modeling.arXiv preprint arXiv:2210.02747 (2022)

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  39. [39]

    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. InProceedings of the IEEE/CVF International Conference on Computer Vision. 14951–14961

    2025

  40. [40]

    Xingchao Liu, Chengyue Gong, and Qiang Liu. 2022. Flow straight and fast: Learning to generate and transfer data with rectified flow.arXiv preprint arXiv:2209.03003(2022)

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  41. [41]

    Yexin Liu, Manyuan Zhang, Yueze Wang, Hongyu Li, Dian Zheng, Weiming Zhang, Changsheng Lu, Xunliang Cai, Yan Feng, Peng Pei, et al. 2025. OpenSub- ject: Leveraging Video-Derived Identity and Diversity Priors for Subject-driven Image Generation and Manipulation.arXiv preprint arXiv:2512.08294(2025)

    work page arXiv 2025

  42. [42]

    Ilya Loshchilov and Frank Hutter. 2017. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101(2017)

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  43. [43]

    Chetwin Low, Weimin Wang, and Calder Katyal. 2025. Ovi: Twin backbone cross-modal fusion for audio-video generation.arXiv preprint arXiv:2510.01284 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  44. [44]

    Ziyang Mai and Yu-Wing Tai. 2025. ContextAnyone: Context-Aware Diffusion for Character-Consistent Text-to-Video Generation.arXiv preprint arXiv:2512.07328 (2025)

    work page arXiv 2025

  45. [45]

    Xiangyu Meng, Zixian Zhang, Zhenghao Zhang, Junchao Liao, Long Qin, and Weizhi Wang. 2025. Identity-grpo: Optimizing multi-human identity-preserving video generation via reinforcement learning.arXiv preprint arXiv:2510.14256 MM ’26, November 10–14, 2026, Rio de Janeiro, Brazil Chen et al. (2025)

    work page arXiv 2025

  46. [46]

    William Peebles and Saining Xie. 2023. Scalable diffusion models with transform- ers. InProceedings of the IEEE/CVF international conference on computer vision. 4195–4205

    2023

  47. [47]

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

    2021

  48. [48]

    Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. 2023. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 22500–22510

    2023

  49. [49]

    Shen Sang, Tiancheng Zhi, Tianpei Gu, Jing Liu, and Linjie Luo. 2025. Lynx: To- wards high-fidelity personalized video generation.arXiv preprint arXiv:2509.15496 (2025)

    work page arXiv 2025

  50. [50]

    Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. 2024. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing568 (2024), 127063

    2024

  51. [51]

    OpenMOSS Team, Donghua Yu, Mingshu Chen, Qi Chen, Qi Luo, Qianyi Wu, Qinyuan Cheng, Ruixiao Li, Tianyi Liang, Wenbo Zhang, et al . 2026. Mova: Towards scalable and synchronized video-audio generation.arXiv preprint arXiv:2602.08794(2026)

    work page arXiv 2026

  52. [52]

    Zachary Teed and Jia Deng. 2020. Raft: Recurrent all-pairs field transforms for optical flow. InEuropean conference on computer vision. Springer, 402–419

    2020

  53. [53]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need.Advances in neural information processing systems30 (2017)

    2017

  54. [54]

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

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  55. [55]

    Yuji Wang, Moran Li, Xiaobin Hu, Ran Yi, Jiangning Zhang, Han Feng, Weijian Cao, Yabiao Wang, Chengjie Wang, and Lizhuang Ma. 2025. Identity-preserving text-to-video generation guided by simple yet effective spatial-temporal decou- pled representations. InProceedings of the 33rd ACM International Conference on Multimedia. 13743–13750

    2025

  56. [56]

    Zhao Wang, Aoxue Li, Lingting Zhu, Yong Guo, Qi Dou, and Zhenguo Li. 2026. Customvideo: Customizing text-to-video generation with multiple subjects.IEEE Transactions on Multimedia(2026)

    2026

  57. [57]

    Jiangchuan Wei, Shiyue Yan, Wenfeng Lin, Boyuan Liu, Renjie Chen, and Mingyu Guo. 2025. Echovideo: Identity-preserving human video generation by multi- modal feature fusion.arXiv preprint arXiv:2501.13452(2025)

    work page arXiv 2025

  58. [58]

    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. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 6537–6549

    2024

  59. [59]

    Bing Wu, Chang Zou, Changlin Li, Duojun Huang, Fang Yang, Hao Tan, Jack Peng, Jianbing Wu, Jiangfeng Xiong, Jie Jiang, et al . 2025. Hunyuanvideo 1.5 technical report.arXiv preprint arXiv:2511.18870(2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  60. [60]

    Tao Wu, Yong Zhang, Xintao Wang, Xianpan Zhou, Guangcong Zheng, Zhongang Qi, Ying Shan, and Xi Li. 2025. Customcrafter: Customized video generation with preserving motion and concept composition abilities. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 39. 8469–8477

    2025

  61. [61]

    Jiahao Xu, Jianjie Luo, and Zhenguo Yang. 2025. Improving Identity Preservation in Video Generation with Multi-Branch Models. InProceedings of the 33rd ACM International Conference on Multimedia. 13758–13765

    2025

  62. [62]

    Xuancheng Xu, Yaning Li, Sisi You, and Bing-Kun Bao. 2025. SMRABooth: Subject and Motion Representation Alignment for Customized Video Generation.arXiv preprint arXiv:2512.12193(2025)

    work page arXiv 2025

  63. [63]

    Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al . 2025. CogVideoX: Text-to-Video Diffusion Models with An Expert Transformer. InThe Thirteenth International Conference on Learning Representations

    2025

  64. [64]

    Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. 2023. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models.arXiv preprint arXiv:2308.06721(2023)

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  65. [65]

    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. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems Datasets and Benchmarks Track

    2025

  66. [66]

    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. InProceedings of the Computer Vision and Pattern Recognition Conference. 12978–12988

    2025

  67. [67]

    Jiangning Zhang, Junwei Zhu, Teng Hu, Yabiao Wang, Donghao Luo, Weijian Cao, Zhenye Gan, Xiaobin Hu, Zhucun Xue, and Chengjie Wang. 2025. Transform Trained Transformer: Accelerating Naive 4K Video Generation Over10 ×.arXiv preprint arXiv:2512.13492(2025)

    work page arXiv 2025

  68. [68]

    Yuechen Zhang, Yaoyang Liu, Bin Xia, Bohao Peng, Zexin Yan, Eric Lo, and Jiaya Jia. 2025. Magicmirror: Id-preserved video generation in video diffusion transformers. InProceedings of the IEEE/CVF International Conference on Computer Vision. 14464–14474

    2025

  69. [69]

    Yiheng Zhang, Zhaofan Qiu, Qi Cai, Yehao Li, Fuchen Long, Yingwei Pan, Ting Yao, and Tao Mei. 2025. Identity-Preserving Video Generation Challenge. In Proceedings of the 33rd ACM International Conference on Multimedia. 13737–13742

    2025

  70. [70]

    Zhenxing Zhang, Jiayan Teng, Zhuoyi Yang, Tiankun Cao, Cheng Wang, Xiaotao Gu, Jie Tang, Dan Guo, and Meng Wang. 2025. Kaleido: Open-Sourced Multi- Subject Reference Video Generation Model.arXiv preprint arXiv:2510.18573 (2025)

    work page arXiv 2025