Pith. sign in

REVIEW 4 major objections 4 minor 81 references

HOMIE claims that injecting pooled multimodal-language-model features directly into a video diffusion transformer's self-attention—while keeping the text encoder—improves both subject fidelity and interaction accuracy in human-object video

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 15:38 UTC pith:37YTYQNR

load-bearing objection Solid architecture and honest ablations, but the SOTA claim is not yet independently verifiable and the paper's own appendix contradicts its headline table. the 4 major comments →

arxiv 2607.18217 v2 pith:37YTYQNR submitted 2026-07-20 cs.CV

HOMIE: Human-object Centric Video Personalization via Multimodal Intelligent Enhancement

Yiyang Cai , Nan Chen , Rongchang Xie , Junwen Pan , Chunyang Jiang , Cheng Chen , Wen Zhou , Zhenbang Sun
show 3 more authors
Wei Xue Wenhan Luo Yike Guo
This is my paper
classification cs.CV
keywords video personalizationhuman-object interactionsubject-driven video generationmultimodal large language modelsself-attention modulationOCR map enhancementmulti-view consistencyflow-matching diffusion transformers
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 is trying to establish that human-object video personalization can be made to work for two different reference structures at once: references that are all different subjects, and references that are multiple views or OCR maps of the same subject. Existing methods either distort the text encoder's control by aligning MLLM features into its embedding space, or they pay a large re-alignment cost by replacing the text encoder entirely. HOMIE keeps the text encoder and adds a third input stream of MLLM-derived features, using Global Multimodal Guidance to inject those features into self-attention and Modality-Reference Embedding to keep token types and identities apart. If the empirical results hold, this would give a cost-effective way to improve subject fidelity, prompt adherence, and text/logo fidelity in product-promotion-style video generation without retraining the whole control pipeline.

Core claim

The central claim is that a video-personalization model can keep its text encoder, feed reference images plus a prompt to a multimodal large-language model, and inject the resulting semantic features into self-attention without a costly re-alignment phase. HOMIE does this with two additions: Global Multimodal Guidance (GMG), which pools the language-model query and key features and applies a feature-wise affine modulation to the video tokens' queries and keys; and Modality-Reference Embedding (MRE), which tags each token by its modality and by which reference entity it belongs to. The paper argues this lets the model connect an abstract logo to the object it belongs on, distinguish several s

What carries the argument

The carrying mechanism is the multimodal input stream F = [video tokens; reference tokens; MLLM tokens] fed through the DiT self-attention. GMG is the injection point: MLLM query and key features are pooled across the temporal dimension, then two lightweight projection networks produce scale and shift factors that modulate the video query and key matrices. MRE is the bookkeeping layer: learnable modality embeddings mark video, reference, and MLLM tokens so cross-modal confusion is reduced, and learnable reference embeddings assign the same vector to references that share a subject identity while using distinct vectors for different subjects. Together they let MLLM knowledge influence video g

Load-bearing premise

The headline advantage depends on a self-curated set of 200 human-object samples whose prompts and similarity judgments are produced by language models, so if those judgments favor what HOMIE was trained on, the reported gap over other methods may not generalize.

What would settle it

Run HOMIE and the strongest existing method on a public, human-object video-personalization benchmark using human-written prompts and human raters for object identity and text fidelity, with baseline models from their released code; if HOMIE's lead on object/logo fidelity and multi-view consistency narrows or disappears, the central claim is beyond what the evidence supports.

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

If this is right

  • A single trained model covers inter-subject and intra-subject personalization, so OCR-map text fidelity and multi-view consistency do not require separate fine-tuning or 3D supervision.
  • Logo personalization is handled by inside-the-model MLLM reasoning: the model can place a logo on the contextually relevant object even when the prompt does not name the relationship.
  • Keeping the text encoder lets the MLLM focus on reference relationships rather than rebuilding text-to-video control, avoiding the re-alignment cost of encoder-replacement designs.
  • Binding multiple references of the same subject with a shared reference embedding prevents those references from being rendered as extra objects in the video.
  • Explicit modality and reference embeddings reduce cross-subject attribute mixing as the number of subjects grows.

Where Pith is reading between the lines

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

  • If GMG works as described, the same 'keep the text encoder, modulate self-attention with pooled multimodal features' pattern is a generic interface that could bring other conditioning signals—depth, audio, camera pose—into a frozen video diffusion model.
  • Because MRE assigns one learned vector per reference entity, a testable extension is to let the number of entities be dynamic; the current fixed-capacity embedding table is the first natural bottleneck to probe.
  • Since the MLLM is frozen and features are extracted offline, the method can likely absorb a stronger MLLM without additional video-model training—an inexpensive upgrading path the paper does not examine.

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 / 4 minor

Summary. The paper proposes HOMIE, a unified framework for human-object centric video personalization (HOCVP) that handles both inter-subject references (distinct humans/objects) and intra-subject references (OCR maps, multi-view images). The method keeps the original text encoder and injects MLLM-derived features through two modules: Global Multimodal Guidance (GMG), which modulates video-token queries/keys with pooled MLLM features inside self-attention, and Modality-Reference Embedding (MRE), which separates tokens by modality and binds intra-subject references. Experiments are conducted on Wan2.1/2.2-14B, with quantitative results reported on a self-curated 200-sample benchmark. The paper claims state-of-the-art performance across HOCVP tasks, with the largest gains in OCR accuracy and multi-view consistency.

Significance. If the quantitative claims are accepted, HOMIE makes a useful architectural contribution: it shows that MLLM knowledge can be injected without replacing or re-aligning the text encoder, and it explicitly addresses intra-subject reference structures that prior methods largely ignore. The ablations in Tab. 3 and Fig. 6 support the individual contributions of GMG and MRE, and the reported training cost (11K A100 hours) is considerably lower than several baselines. However, the central 'state-of-the-art' claim rests on an unreleased, self-curated evaluation set and LLM-as-judge metrics that are not validated against standard tools. One appendix passage explicitly states that two comparison methods achieve stronger performance than HOMIE, which directly undercuts the main claim. The strengths of the method design are real, but the evidence as presented is not yet sufficient to establish the headline result.

major comments (4)
  1. [Appendix A.2.1] The manuscript contains an internal contradiction that is load-bearing for the central claim. After arguing that UniVideo and VINO replace the text encoder and incur higher cost, A.2.1 states that these methods 'achieve stronger performance than their Wan-14B-based counterparts, including HOMIE.' This directly contradicts the abstract and Tab. 1, where HOMIE is presented as state-of-the-art and UniVideo/VINO report lower scores. If the sentence is meant to say something else (e.g., stronger than their Wan-based counterparts but not than HOMIE), it must be corrected; as written, the paper's own appendix undermines the SOTA assertion.
  2. [Appendix A.2.2/A.2.3, Tab. 1] The largest reported gain, OCR accuracy (0.452 vs 0.326 for SkyReels-V3), is measured by GPT-5.2 with no validation against a standard OCR engine or human annotation. Since HOMIE is explicitly trained with OCR-map intra-subject references, the judge may reward plausible text-like rendering rather than exact string match. Eq. (12) defines normalized Levenshtein similarity, but the inputs to that formula are GPT-5.2 extractions; no evidence is given that GPT-5.2's extraction is accurate enough to support a 21.8% relative improvement claim. The evaluation set is also self-curated (200 samples) and its release is deferred to 'later,' so the central quantitative comparison is not independently reproducible.
  3. [Tab. 1, Tab. 7, Section 3] Baseline comparability is not established. Tab. 1 mixes closed-source Kling, Wan2.1/2.2-based open methods, and HunyuanVideo-based UniVideo/VINO with different training pipelines. The claim that HOMIE's integration strategy is 'better' than text-encoder replacement requires controlled experiments on the same backbone with comparable data and compute. The ablation 'MLLM to UmT5' (Tab. 3d) is conducted on Wan with an identical alignment network, which is informative, but it does not evaluate the full replace-the-text-encoder architecture on the same backbone. Without this control, the observed gains could be due to backbone, training recipe, or compute rather than the proposed GMG/MRE modules.
  4. [Section 4, Tabs. 1-3, Fig. 7] No confidence intervals, error bars, or significance tests are reported for any quantitative metric. On a 200-sample self-curated set, several differences are small in absolute terms (e.g., DINO-I 0.543 vs 0.526, Obj-Sim 0.891 vs 0.872, DINOacc 0.582 vs 0.560 in Tab. 2). The user study (40 participants, 20 samples) also reports only preference percentages without statistical testing. Given that the evaluation set is not public, the absence of uncertainty quantification makes it impossible to assess whether the reported margins are meaningful or within sampling noise.
minor comments (4)
  1. [Eq. (3)-(9), Sec. 3.1.3] The notation for token concatenation and the 'Proj' function in MRE is a bit underspecified. Please define the projection function precisely, e.g., how each modality index maps to slice indices, and clarify whether video tokens receive both a modality embedding and a reference embedding.
  2. [Appendix A.2.3] The OCR extraction prompt and the object similarity prompt are given, but no examples of GPT-5.2's outputs or failure modes are provided. Please report agreement with a conventional OCR engine (e.g., PaddleOCR or Tesseract) on the same sampled frames.
  3. [Tab. 3, Fig. 6] The ablation setting 'MLLM to UmT5' is described only briefly. Please state explicitly whether GMG is removed in this setting or whether the MLLM features are simply appended to the UmT5 embedding sequence while GMG remains active; the qualitative comparison in Fig. 6(a) suggests a full integration change, but the exact configuration should be unambiguous.
  4. [Throughout] Minor typographical and formatting issues appear, e.g., 'V AE' with odd spacing in several equations and table captions, and the inconsistent use of 'V AE tokens' vs 'VAE tokens.' These do not affect the technical content.

Circularity Check

0 steps flagged

No by-construction circularity: GMG/MRE are trained input-conditioning modules; the SOTA claim rests on evaluation validity, not on a circular derivation.

full rationale

Walking the derivation chain: the method is an input-conditioning architecture. Eq. (1) defines a flow-matching training objective against ground-truth velocity; Eqs. (2)-(9) define how reference images and MLLM hidden states are mapped into tokens and modulated inside self-attention. None of these equations defines an evaluation target (OCR accuracy, DINO-I, Obj-Sim, Face-Sim) as a function of the fitted parameters. GMG and MRE are trained network components, not post-hoc fits, and the reported numbers in Tabs. 1-3 are outputs of the trained model scored by external or semi-external metrics. No reported 'prediction' is the value of a fitted parameter renamed as a result. The self-citations in the manuscript ([5], [10], [31], [72]) are used as general related-work context for personalization, animation, and editing; they are not invoked to justify HOMIE's design, to supply a uniqueness theorem, or to forbid alternative architectures. There is also no ansatz smuggled in via a self-citation: the MLLM integration strategy and the GMG/MRE modules are introduced in the paper itself rather than imported from the authors' prior work. The strongest skeptical concern is real but is an evaluation-validity issue, not circularity: the 200-sample self-curated benchmark and the GPT-5.2-based OCR/object-similarity judge described in Appendix A.2.2-A.2.3 could be biased toward HOMIE's training distribution, and the appendices even contain the sentence that UniVideo/VINO 'achieve stronger performance than their Wan-14B-based counterparts, including HOMIE,' which undercuts the abstract's SOTA claim. Such issues bear on whether the reported metric gaps generalize, not on whether any result is equivalent to its inputs by construction. Therefore no significant circularity is present.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

The central claim rests on transfer of MLLM relational knowledge into an existing video diffusion transformer and on the validity of LLM-judged metrics. GMG and MRE are learned modules rather than new physical entities or post-hoc fitted constants. The main hand-chosen free parameters are the reference-embedding count, the MLLM token cap, the training schedule, and unspecified dataset-filtering thresholds.

free parameters (4)
  • number of MRE reference embeddings = 5
    Set to 5 in implementation (§4); bounds how many distinct subjects/identities can be tagged, with no sensitivity analysis reported.
  • MLLM feature sequence length cap = 1024 tokens
    Output hidden states capped at 1024 tokens (§A.1.1); affects how much MLLM information reaches the DiT, no ablation.
  • three-stage training schedule = 3000/2000/500 steps, LR 1e-5/1e-5/2e-6
    Hand-tuned schedule in Tab. 5; no systematic search reported.
  • dataset filtering thresholds = unspecified
    Aesthetic scores and area proportions are used to filter OpenS2V-5M and PhantomData (§3.2), but exact thresholds are not given, making the training distribution hard to replicate.
axioms (5)
  • domain assumption Pre-trained Wan2.1/2.2 T2V model provides stable text control and a good video prior under the flow-matching objective (Eq. 1).
    The method builds on Wan and assumes its text control survives when MLLM tokens are concatenated into self-attention.
  • domain assumption Qwen3-VL-2B hidden states encode inter- and intra-subject relationships in a form injectable into video self-attention through a 2-layer MLP.
    This is the core transfer assumption behind GMG; the paper provides empirical support but no mechanistic guarantee.
  • domain assumption GPT-5.2-based object similarity and OCR accuracy correlate with true subject fidelity.
    Automatic evaluation uses an LLM as judge (Tab. 8), so metric validity is assumed rather than calibrated against human ratings beyond one 40-person user study.
  • domain assumption The self-curated 200-sample evaluation set is representative of HOCVP tasks.
    The split (100 inter-subject, 60 OCR, 40 multi-view) and sample composition are chosen by the authors; no public benchmark is used.
  • domain assumption Reference annotations in OpenS2V-5M and PhantomData are reliable enough for filtering HOC-relevant subjects and clips.
    Dataset curation relies on category labels and area proportions provided by the source datasets (§3.2).

pith-pipeline@v1.3.0-alltime-deepseek · 20298 in / 10555 out tokens · 83690 ms · 2026-08-01T15:38:26.935170+00:00 · methodology

0 comments
read the original abstract

Human-object centric video personalization (HOCVP) is a core task within subject-driven video generation. However, existing methods suffer from two key limitations. First, most approaches focusing on inter-subject personalization still struggle to strike a balance between high subject fidelity and accurate interaction patterns between humans and diverse objects, especially when objects represent abstract concepts such as logos. Second, while intra-subject references (e.g., OCR maps, multi-view inputs) are expected to enhance subject fidelity, most existing works lack mechanisms to understand such latent correspondence. To address both challenges, we propose HOMIE, an HOCVP framework that tackles both inter- and intra-subject input settings in a unified manner. Compared to previous approaches, HOMIE proposes a better MLLM integration strategy to extract knowledge of reference-level relationships without compromising the controllability of text encoders or incurring costly re-alignment. Specifically, we introduce global multimodal guidance within self-attention to better align MLLM-derived semantic features with VAE tokens. Furthermore, we propose modality-reference embedding to differentiate tokens from MLLM features and VAE tokens and associate intra-subject reference image tokens. Extensive experiments validate that our method achieves state-of-the-art performance across various HOCVP tasks. Project Page: https://yiyangcai.github.io/homie-page.github.io/

Figures

Figures reproduced from arXiv: 2607.18217 by Cheng Chen, Chunyang Jiang, Junwen Pan, Nan Chen, Rongchang Xie, Wei Xue, Wenhan Luo, Wen Zhou, Yike Guo, Yiyang Cai, Zhenbang Sun.

Figure 1
Figure 1. Figure 1: HOMIE addresses HOCVP with both inter/intra-subject references: (1) multi-human [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Framework design of MLLM-facilitated video personalization. [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overview of HOMIE. HOMIE employs a multimodal-input paradigm with video tokens, image tokens, and multimodal features. Prior to self-attention, Global Multimodal Guidance (GMG) fuses high-semantic MLLM features into video tokens, boosting cross-modal interaction. Furthermore, Multimodal-Reference Embedding (MRE) enables the model to discriminate cross-modal tokens and inter-subject tokens, as well as bind … view at source ↗
Figure 4
Figure 4. Figure 4: Inter-subject qualitative comparison between HOMIE and previous SOTA methods, including more subjects and abstract content personalization. Prompts that are relevant to subjects and interaction are highlighted. (Zoom in for the best view) its advantage compared with architectures replacing UmT5 with MLLMs (Fig.2(b)), we additionally evaluate against UniVideo[63] and VINO[8], which utilize the HunyuanVideo-… view at source ↗
Figure 5
Figure 5. Figure 5: Intra-subject qualitative comparison between HOMIE and previous SOTA methods. Two representative intra-subject exemplars (OCR maps and multi-view references) are provided. Prompts that are relevant to subjects and interaction are highlighted. (Zoom in for the best view) SkyReels-V3), validating its effective use of OCR maps from the evaluation dataset. Moreover, as reported in Tab. 2, HOMIE also delivers c… view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative Ablation Studies (Zoom in for the best view.) “MLLM to UmT5” denotes the integration of MLLM features into the UmT5 embeddings, which corresponds to the architecture illustrated in [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: User study results. Modality-Reference Embedding [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: More inter-subject qualitative comparison between HOMIE and previous SOTA methods, including more subjects and complex interaction patterns. Prompts that are relevant to subjects and interaction are highlighted. (Zoom in for the best view) 21 [PITH_FULL_IMAGE:figures/full_fig_p021_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: More inter-subject qualitative comparison between HOMIE and previous SOTA methods, including more subjects and complex interaction patterns. Prompts that are relevant to subjects and interaction are highlighted. (Zoom in for the best view) 22 [PITH_FULL_IMAGE:figures/full_fig_p022_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: More qualitative comparison between HOMIE and previous SOTA methods (integrated [PITH_FULL_IMAGE:figures/full_fig_p023_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: More intra-subject qualitative comparison between HOMIE and previous SOTA methods. Two representative intra-subject exemplars (OCR maps and multi-view references) are provided. Prompts that are relevant to subjects and interaction are highlighted. (Zoom in for the best view) 24 [PITH_FULL_IMAGE:figures/full_fig_p024_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Additional visual results of HOMIE on inter-subject HOCVP scenarios, including general [PITH_FULL_IMAGE:figures/full_fig_p026_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Additional visual results of HOMIE on intra-subject HOCVP scenarios, including OCR [PITH_FULL_IMAGE:figures/full_fig_p027_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Demonstration of real application (e.g., product promotion) [PITH_FULL_IMAGE:figures/full_fig_p028_14.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

81 extracted references · 25 linked inside Pith

  1. [1]

    You are an expert multimodal AI assistant specialized in highly controllable visual content generation

  2. [2]

    Dynamic concepts personalization from single videos

    Rameen Abdal, Or Patashnik, Ivan Skorokhodov, Willi Menapace, Aliaksandr Siarohin, Sergey Tulyakov, Daniel Cohen-Or, and Kfir Aberman. Dynamic concepts personalization from single videos. InProceedings of the Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Papers, pages 1–9, 2025

    2025

  3. [3]

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

    Pith/arXiv arXiv 2025

  4. [4]

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

    Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Do- minik 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, 2023

    Pith/arXiv arXiv 2023

  5. [5]

    Yiyang Cai, Zhengkai Jiang, Yulong Liu, Chunyang Jiang, Wei Xue, Yike Guo, and Wenhan Luo. Foundation cures personalization: Improving personalized models’ prompt consistency via hidden foundation knowledge.Advances in Neural Information Processing Systems, 38:12776– 12814, 2026

    2026

  6. [6]

    Omnivcus: Feedforward subject-driven video customization with multimodal control conditions

    Yuanhao Cai, He Zhang, Xi Chen, Jinbo Xing, Yiwei Hu, Yuqian Zhou, Kai Zhang, Zhifei Zhang, Soo Ye Kim, Tianyu Wang, et al. Omnivcus: Feedforward subject-driven video customization with multimodal control conditions. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025

    2025

  7. [7]

    First frame is the place to go for video content customization.arXiv preprint arXiv:2511.15700, 2025

    Jingxi Chen, Zongxia Li, Zhichao Liu, Guangyao Shi, Xiyang Wu, Fuxiao Liu, Cornelia Fermuller, Brandon Y Feng, and Yiannis Aloimonos. First frame is the place to go for video content customization.arXiv preprint arXiv:2511.15700, 2025

    arXiv 2025

  8. [8]

    Vino: A unified visual generator with interleaved omnimodal context.arXiv preprint arXiv:2601.02358, 2026

    Junyi Chen, Tong He, Zhoujie Fu, Pengfei Wan, Kun Gai, and Weicai Ye. Vino: A unified visual generator with interleaved omnimodal context.arXiv preprint arXiv:2601.02358, 2026

    arXiv 2026

  9. [9]

    Human-centric video generation via collaborative multi-modal conditioning

    Liyang Chen, Tianxiang Ma, Jiawei Liu, Bingchuan Li, Zhuowei Chen, Lijie Liu, Xu He, Gen Li, Qian He, and Zhiyong Wu. Human-centric video generation via collaborative multi-modal conditioning. InProceedings of the AAAI Conference on Artificial Intelligence, volume 40, pages 2939–2947, 2026

    2026

  10. [10]

    Domainshuttle: Freeform open domain subject-driven text-to-video generation.arXiv preprint arXiv:2606.26058, 2026

    Nan Chen, Yiyang Cai, Rongchang Xie, Junwen Pan, Cheng Chen, Weinan Jia, Zhuowei Chen, Wen Zhou, Zhenbang Sun, and Wenhan Luo. Domainshuttle: Freeform open domain subject-driven text-to-video generation.arXiv preprint arXiv:2606.26058, 2026

    Pith/arXiv arXiv 2026

  11. [11]

    Multi-subject open-set personalization in video generation

    Tsai-Shien Chen, Aliaksandr Siarohin, Willi Menapace, Yuwei Fang, Kwot Sin Lee, Ivan Sko- rokhodov, Kfir Aberman, Jun-Yan Zhu, Ming-Hsuan Yang, and Sergey Tulyakov. Multi-subject open-set personalization in video generation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6099–6110, June 2025

    2025

  12. [12]

    Phantom-data: Towards a general subject-consistent video generation dataset

    Zhuowei Chen, Bingchuan Li, Tianxiang Ma, Lijie Liu, Mingcong Liu, Yunsheng Jiang, Gen Li, Xinghui Li, Liyang Chen, SiYu Zhou, Qian HE, and Xinglong Wu. Phantom-data: Towards a general subject-consistent video generation dataset. InThe Fourteenth International Conference on Learning Representations, 2026. 10

    2026

  13. [13]

    Unimax: Fairer and more effective language sampling for large-scale multilingual pretraining.arXiv preprint arXiv:2304.09151, 2023

    Hyung Won Chung, Noah Constant, Xavier Garcia, Adam Roberts, Yi Tay, Sharan Narang, and Orhan Firat. Unimax: Fairer and more effective language sampling for large-scale multilingual pretraining.arXiv preprint arXiv:2304.09151, 2023

    Pith/arXiv arXiv 2023

  14. [14]

    Cinema: Coherent multi-subject video generation via mllm-based guidance.arXiv preprint arXiv:2503.10391, 2025

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

    Pith/arXiv arXiv 2025

  15. [15]

    MAGREF: Masked guidance for any-reference video generation with subject disentanglement

    Yufan Deng, Yuanyang Yin, Xun Guo, Yizhi Wang, Jacob Zhiyuan Fang, Shenghai Yuan, Yiding Yang, Angtian Wang, Bo Liu, Haibin Huang, and Chongyang Ma. MAGREF: Masked guidance for any-reference video generation with subject disentanglement. InThe Fourteenth International Conference on Learning Representations, 2026

    2026

  16. [16]

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

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

    Pith/arXiv arXiv 2025

  17. [17]

    Identity-preserving text-to-video generation via training-free prompt, image, and guidance enhancement

    Jiayi Gao, Changcheng Hua, Qingchao Chen, Yuxin Peng, and Yang Liu. Identity-preserving text-to-video generation via training-free prompt, image, and guidance enhancement. InProceed- ings of the 33rd ACM International Conference on Multimedia, MM ’25, page 13751–13757, New York, NY , USA, 2025. Association for Computing Machinery

    2025

  18. [18]

    Wildactor: Unconstrained identity-preserving video generation.arXiv preprint arXiv:2603.00586, 2026

    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, 2026

    arXiv 2026

  19. [19]

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

    Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. InThe Twelfth International Conference on Learning Representations, 2024

    2024

  20. [20]

    Pulid: Pure and lightning id customization via contrastive alignment.Advances in neural information processing systems, 37:36777–36804, 2024

    Zinan Guo, Yanze Wu, Zhuowei Chen, Lang Chen, Peng Zhang, and Qian He. Pulid: Pure and lightning id customization via contrastive alignment.Advances in neural information processing systems, 37:36777–36804, 2024

    2024

  21. [21]

    Ltx-video: Realtime video latent diffusion.arXiv preprint arXiv:2501.00103, 2024

    Yoav HaCohen, Nisan Chiprut, Benny Brazowski, Daniel Shalem, Dudu Moshe, Eitan Richard- son, Eran Levin, Guy Shiran, Nir Zabari, Ori Gordon, et al. Ltx-video: Realtime video latent diffusion.arXiv preprint arXiv:2501.00103, 2024

    Pith/arXiv arXiv 2024

  22. [22]

    Id-animator: Zero-shot identity-preserving human video generation.arXiv preprint arXiv:2404.15275, 2024

    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, 2024

    Pith/arXiv arXiv 2024

  23. [23]

    Video diffusion models.Advances in neural information processing systems, 35:8633–8646, 2022

    Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models.Advances in neural information processing systems, 35:8633–8646, 2022

    2022

  24. [24]

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

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

    Pith/arXiv arXiv 2025

  25. [25]

    Polyvivid: Vivid multi-subject video generation with cross-modal interaction and enhancement

    Teng Hu, Zhentao Yu, Zhengguang Zhou, Jiangning Zhang, Yuan Zhou, Qinglin Lu, and Ran Yi. Polyvivid: Vivid multi-subject video generation with cross-modal interaction and enhancement. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025

    2025

  26. [26]

    Adahuman: Animatable detailed 3d human generation with compositional multiview diffusion

    Yangyi Huang, Ye Yuan, Xueting Li, Jan Kautz, and Umar Iqbal. Adahuman: Animatable detailed 3d human generation with compositional multiview diffusion. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 13533–13543, 2025

    2025

  27. [27]

    Conceptmaster: Multi-concept video customization on diffusion transformer models without test-time tuning.arXiv preprint arXiv:2501.04698, 2025

    Yuzhou Huang, Ziyang Yuan, Quande Liu, Qiulin Wang, Xintao Wang, Ruimao Zhang, Pengfei Wan, Di Zhang, and Kun Gai. Conceptmaster: Multi-concept video customization on diffusion transformer models without test-time tuning.arXiv preprint arXiv:2501.04698, 2025. 11

    Pith/arXiv arXiv 2025

  28. [28]

    Vace: All-in- one video creation and editing

    Zeyinzi Jiang, Zhen Han, Chaojie Mao, Jingfeng Zhang, Yulin Pan, and Yu Liu. Vace: All-in- one video creation and editing. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 17191–17202, 2025

    2025

  29. [29]

    Kling api.https://klingai.com/global/, 2025

    Kling. Kling api.https://klingai.com/global/, 2025. Accessed: 2025-01-01

    2025

  30. [30]

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

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

    Pith/arXiv arXiv 2024

  31. [31]

    Let them talk: Audio-driven multi-person conversational video generation

    Zhe Kong, Feng Gao, Yong Zhang, Zhuoliang Kang, Xiaoming Wei, Xunliang Cai, Guanying Chen, and Wenhan Luo. Let them talk: Audio-driven multi-person conversational video generation. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems, 2026

    2026

  32. [32]

    Skyreels-v3 technique report.arXiv preprint arXiv:2601.17323, 2026

    Debang Li, Zhengcong Fei, Tuanhui Li, Yikun Dou, Zheng Chen, Jiangping Yang, Mingyuan Fan, Jingtao Xu, Jiahua Wang, Baoxuan Gu, et al. Skyreels-v3 technique report.arXiv preprint arXiv:2601.17323, 2026

    arXiv 2026

  33. [33]

    Magicid: Hybrid preference optimization for id-consistent and dynamic-preserved video cus- tomization

    Hengjia Li, Lifan Jiang, Xi Xiao, Tianyang Wang, Hongwei Yi, Boxi Wu, and Deng Cai. Magicid: Hybrid preference optimization for id-consistent and dynamic-preserved video cus- tomization. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12737–12746, October 2025

    2025

  34. [34]

    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 (ICCV), pages 19406–19416, October 2025

    2025

  35. [35]

    Bindweave: Subject-consistent video generation via cross- modal integration

    Zhaoyang Li, Dongjun Qian, Kai Su, qishuai diao, Xiangyang Xia, Chang Liu, Wenfei Yang, Tianzhu Zhang, and Zehuan Yuan. Bindweave: Subject-consistent video generation via cross- modal integration. InThe Fourteenth International Conference on Learning Representations, 2026

    2026

  36. [36]

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

    Feng Liang, Haoyu Ma, Zecheng He, Tingbo Hou, Ji Hou, Kunpeng Li, Xiaoliang Dai, Felix Juefei-Xu, Samaneh Azadi, Animesh Sinha, et al. Movie weaver: Tuning-free multi-concept video personalization with anchored prompts. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 13146–13156, 2025

    2025

  37. [37]

    Open-sora plan: Open-source large video generation model.arXiv preprint arXiv:2412.00131, 2024

    Bin Lin, Yunyang Ge, Xinhua Cheng, Zongjian Li, Bin Zhu, Shaodong Wang, Xianyi He, Yang Ye, Shenghai Yuan, Liuhan Chen, et al. Open-sora plan: Open-source large video generation model.arXiv preprint arXiv:2412.00131, 2024

    Pith/arXiv arXiv 2024

  38. [38]

    Uniworld: High-resolution semantic encoders for unified visual understanding and generation.arXiv preprint arXiv:2506.03147, 2025

    Bin Lin, Zongjian Li, Xinhua Cheng, Yuwei Niu, Yang Ye, Xianyi He, Shenghai Yuan, Wangbo Yu, Shaodong Wang, Yunyang Ge, et al. Uniworld: High-resolution semantic encoders for unified visual understanding and generation.arXiv preprint arXiv:2506.03147, 2025

    Pith/arXiv arXiv 2025

  39. [39]

    Exploring mllm-diffusion information transfer with metacanvas.arXiv preprint arXiv:2512.11464, 2025

    Han Lin, Xichen Pan, Ziqi Huang, Ji Hou, Jialiang Wang, Weifeng Chen, Zecheng He, Felix Juefei-Xu, Junzhe Sun, Zhipeng Fan, et al. Exploring mllm-diffusion information transfer with metacanvas.arXiv preprint arXiv:2512.11464, 2025

    arXiv 2025

  40. [40]

    Mofu: Scale-aware modulation and fourier fusion for multi- subject video generation

    Run Ling, Ke Cao, Jian Lu, Ao Ma, Haowei Liu, Runze He, Changwei Wang, Rongtao Xu, Yihua Shao, Zhanjie Zhang, et al. Mofu: Scale-aware modulation and fourier fusion for multi- subject video generation. InProceedings of the AAAI Conference on Artificial Intelligence, volume 40, pages 7033–7041, 2026

    2026

  41. [41]

    Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matthew Le. Flow matching for generative modeling. InThe Eleventh International Conference on Learning Representations, 2023

    2023

  42. [42]

    Visual instruction tuning.Advances in neural information processing systems, 36:34892–34916, 2023

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning.Advances in neural information processing systems, 36:34892–34916, 2023. 12

    2023

  43. [43]

    Phantom: Subject-consistent video generation via cross-modal alignment

    Lijie Liu, Tianxiang Ma, Bingchuan Li, Zhuowei Chen, Jiawei Liu, Gen Li, Siyu Zhou, Qian He, and Xinglong Wu. Phantom: Subject-consistent video generation via cross-modal alignment. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14951–14961, October 2025

    2025

  44. [44]

    Step1x-edit: A practical framework for general image editing.arXiv preprint arXiv:2504.17761, 2025

    Shiyu Liu, Yucheng Han, Peng Xing, Fukun Yin, Rui Wang, Wei Cheng, Jiaqi Liao, Yingming Wang, Honghao Fu, Chunrui Han, et al. Step1x-edit: A practical framework for general image editing.arXiv preprint arXiv:2504.17761, 2025

    Pith/arXiv arXiv 2025

  45. [45]

    Contextanyone: Context-aware diffusion for character-consistent text-to-video generation.arXiv preprint arXiv:2512.07328, 2025

    Ziyang Mai and Yu-Wing Tai. Contextanyone: Context-aware diffusion for character-consistent text-to-video generation.arXiv preprint arXiv:2512.07328, 2025

    arXiv 2025

  46. [46]

    Instructx: Towards unified visual editing with mllm guidance.arXiv preprint arXiv:2510.08485, 2025

    Chong Mou, Qichao Sun, Yanze Wu, Pengze Zhang, Xinghui Li, Fulong Ye, Songtao Zhao, and Qian He. Instructx: Towards unified visual editing with mllm guidance.arXiv preprint arXiv:2510.08485, 2025

    arXiv 2025

  47. [47]

    OpenAI. Gpt-5.2. https://platform.openai.com/docs/models/gpt-5.2, 2025. Ac- cessed: 2025-01-01

    2025

  48. [48]

    Dinov2: Learning robust visual features without supervision.arXiv preprint arXiv:2304.07193, 2023

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

    Pith/arXiv arXiv 2023

  49. [49]

    Omniweaving: Towards unified video generation with free-form composition and reasoning.arXiv preprint arXiv:2603.24458, 2026

    Kaihang Pan, Qi Tian, Jianwei Zhang, Weijie Kong, Jiangfeng Xiong, Yanxin Long, Shixue Zhang, Haiyi Qiu, Tan Wang, Zheqi Lv, et al. Omniweaving: Towards unified video generation with free-form composition and reasoning.arXiv preprint arXiv:2603.24458, 2026

    arXiv 2026

  50. [50]

    Id-crafter: Vlm-grounded online rl for compositional multi-subject video generation.arXiv preprint arXiv:2511.00511, 2025

    Panwang Pan, Jingjing Zhao, Yuchen Lin, Chenguo Lin, Chenxin Li, Hengyu Liu, Tingting Shen, and Yadong Mu. Id-crafter: Vlm-grounded online rl for compositional multi-subject video generation.arXiv preprint arXiv:2511.00511, 2025

    arXiv 2025

  51. [51]

    Transfer between modalities with metaqueries.arXiv preprint arXiv:2504.06256, 2025

    Xichen Pan, Satya Narayan Shukla, Aashu Singh, Zhuokai Zhao, Shlok Kumar Mishra, Jialiang Wang, Zhiyang Xu, Jiuhai Chen, Kunpeng Li, Felix Juefei-Xu, et al. Transfer between modalities with metaqueries.arXiv preprint arXiv:2504.06256, 2025

    Pith/arXiv arXiv 2025

  52. [52]

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

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

    Pith/arXiv arXiv 2024

  53. [53]

    Fatezero: Fusing attentions for zero-shot text-based video editing

    Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 15932–15942, 2023

    2023

  54. [54]

    High- resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High- resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022

    2022

  55. [55]

    Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation

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

    2023

  56. [56]

    Lynx: Towards high-fidelity personalized video generation.arXiv preprint arXiv:2509.15496, 2025

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

    arXiv 2025

  57. [57]

    Seedance 1.5 pro: A native audio-visual joint generation foundation model.arXiv preprint arXiv:2512.13507, 2025

    Team Seedance, Heyi Chen, Siyan Chen, Xin Chen, Yanfei Chen, Ying Chen, Zhuo Chen, Feng Cheng, Tianheng Cheng, Xinqi Cheng, et al. Seedance 1.5 pro: A native audio-visual joint generation foundation model.arXiv preprint arXiv:2512.13507, 2025

    Pith/arXiv arXiv 2025

  58. [58]

    Mv-s2v: Multi-view subject-consistent video generation (2026)

    Z Song, X Gong, B Liu, and Z Zhao. Mv-s2v: Multi-view subject-consistent video generation (2026). 13

    2026

  59. [59]

    Mvhoi: Bridge multi-view condition to complex human-object interaction video reenactment via 3d foundation model.arXiv preprint arXiv:2603.14686, 2026

    Jinguang Tong, Jinbo Wu, Kaisiyuan Wang, Zhelun Shen, Xuan Huang, Mochu Xiang, Xuesong Li, Yingying Li, Haocheng Feng, Chen Zhao, et al. Mvhoi: Bridge multi-view condition to complex human-object interaction video reenactment via 3d foundation model.arXiv preprint arXiv:2603.14686, 2026

    arXiv 2026

  60. [60]

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

    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, 2025

    Pith/arXiv arXiv 2025

  61. [61]

    Refalign: Representation alignment for reference-to-video generation.arXiv preprint arXiv:2603.25743, 2026

    Lei Wang, YuXin Song, Ge Wu, Haocheng Feng, Hang Zhou, Jingdong Wang, Yaxing Wang, et al. Refalign: Representation alignment for reference-to-video generation.arXiv preprint arXiv:2603.25743, 2026

    Pith/arXiv arXiv 2026

  62. [62]

    Dreamactor-h1: High-fidelity human-product demonstration video generation via motion-designed diffusion transformers.arXiv preprint arXiv:2506.10568, 2025

    Lizhen Wang, Zhurong Xia, Tianshu Hu, Pengrui Wang, Pengfei Wei, Zerong Zheng, Ming Zhou, Yuan Zhang, and Mingyuan Gao. Dreamactor-h1: High-fidelity human-product demonstration video generation via motion-designed diffusion transformers.arXiv preprint arXiv:2506.10568, 2025

    Pith/arXiv arXiv 2025

  63. [63]

    Univideo: Unified understanding, generation, and editing for videos

    Cong Wei, Quande Liu, Zixuan Ye, Qiulin Wang, Xintao Wang, Pengfei Wan, Kun Gai, and Wenhu Chen. Univideo: Unified understanding, generation, and editing for videos. InThe Fourteenth International Conference on Learning Representations, 2026

    2026

  64. [64]

    Dreamvideo: Composing your dream videos with cus- tomized subject and motion

    Yujie Wei, Shiwei Zhang, Zhiwu Qing, Hangjie Yuan, Zhiheng Liu, Yu Liu, Yingya Zhang, Jingren Zhou, and Hongming Shan. Dreamvideo: Composing your dream videos with cus- tomized subject and motion. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6537–6549, 2024

    2024

  65. [65]

    Omnigen2: Exploration to advanced multimodal generation.arXiv preprint arXiv:2506.18871, 2025

    Chenyuan Wu, Pengfei Zheng, Ruiran Yan, Shitao Xiao, Xin Luo, Yueze Wang, Wanli Li, Xiyan Jiang, Yexin Liu, Junjie Zhou, et al. Omnigen2: Exploration to advanced multimodal generation.arXiv preprint arXiv:2506.18871, 2025

    Pith/arXiv arXiv 2025

  66. [66]

    Consid-gen: View-consistent and identity-preserving image-to-video generation.arXiv preprint arXiv:2602.10113, 2026

    Mingyang Wu, Ashirbad Mishra, Soumik Dey, Shuo Xing, Naveen Ravipati, Hansi Wu, Binbin Li, and Zhengzhong Tu. Consid-gen: View-consistent and identity-preserving image-to-video generation.arXiv preprint arXiv:2602.10113, 2026

    arXiv 2026

  67. [67]

    Structured 3d latents for scalable and versatile 3d generation

    Jianfeng Xiang, Zelong Lv, Sicheng Xu, Yu Deng, Ruicheng Wang, Bowen Zhang, Dong Chen, Xin Tong, and Jiaolong Yang. Structured 3d latents for scalable and versatile 3d generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21469–21480, 2025

    2025

  68. [68]

    Lumosx: Relate any identities with their attributes for personalized video generation

    Jiazheng Xing, Fei Du, Hangjie Yuan, Pengwei Liu, Hongbin Xu, Hai Ci, Ruigang Niu, Weihua Chen, Fan Wang, and Yong Liu. Lumosx: Relate any identities with their attributes for personalized video generation. InThe Fourteenth International Conference on Learning Representations, 2026

    2026

  69. [69]

    Anchorcrafter: Animate cyberanchors saling your products via human-object interacting video generation.arXiv preprint arXiv:2411.17383, 2024

    Ziyi Xu, Ziyao Huang, Juan Cao, Yong Zhang, Xiaodong Cun, Qing Shuai, Yuchen Wang, Linchao Bao, Jintao Li, and Fan Tang. Anchorcrafter: Animate cyberanchors saling your products via human-object interacting video generation.arXiv preprint arXiv:2411.17383, 2024

    Pith/arXiv arXiv 2024

  70. [70]

    Stand-in: A lightweight and plug-and-play identity control for video generation.arXiv preprint arXiv:2508.07901, 2025

    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, 2025

    arXiv 2025

  71. [71]

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

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

    Pith/arXiv arXiv 2024

  72. [72]

    Unified in-context video editing

    Zixuan Ye, Xuanhua He, Quande Liu, Qiulin Wang, Xintao Wang, Pengfei Wan, Di ZHANG, Kun Gai, Qifeng Chen, and Wenhan Luo. Unified in-context video editing. InThe Fourteenth International Conference on Learning Representations, 2026. 14

    2026

  73. [73]

    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, Chongyang Ma, Jiebo Luo, and Li Yuan. 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, 2026

    2026

  74. [74]

    Identity-preserving text-to-video generation by frequency decomposition

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

    2025

  75. [75]

    Proteus-id: Id-consistent and motion-coherent video customization

    Guiyu Zhang, Chen Shi, Zijian Jiang, Xunzhi Xiang, Jingjing Qian, Shaoshuai Shi, and Li Jiang. Proteus-id: Id-consistent and motion-coherent video customization. InProceedings of the SIGGRAPH Asia 2025 Conference Papers, SA Conference Papers ’25, New York, NY , USA,

    2025

  76. [76]

    Gme: Improving universal multimodal retrieval by multimodal llms.arXiv preprint arXiv:2412.16855, 2024

    Xin Zhang, Yanzhao Zhang, Wen Xie, Mingxin Li, Ziqi Dai, Dingkun Long, Pengjun Xie, Meishan Zhang, Wenjie Li, and Min Zhang. Gme: Improving universal multimodal retrieval by multimodal llms.arXiv preprint arXiv:2412.16855, 2024

    Pith/arXiv arXiv 2024

  77. [77]

    Anytalker: Scaling multi-person talking video generation with interactivity refinement.arXiv preprint arXiv:2511.23475, 2025

    Zhizhou Zhong, Yicheng Ji, Zhe Kong, Yiying Liu, Jiarui Wang, Jiasun Feng, Lupeng Liu, Xiangyi Wang, Yanjia Li, Yuqing She, et al. Anytalker: Scaling multi-person talking video generation with interactivity refinement.arXiv preprint arXiv:2511.23475, 2025

    arXiv 2025

  78. [78]

    Brandfusion: A multi- agent framework for seamless brand integration in text-to-video generation.arXiv preprint arXiv:2603.02816, 2026

    Zihao Zhu, Ruotong Wang, Siwei Lyu, Min Zhang, and Baoyuan Wu. Brandfusion: A multi- agent framework for seamless brand integration in text-to-video generation.arXiv preprint arXiv:2603.02816, 2026. 15 A Appendix Our supplementary material is organized into the following sections: • Section A.1 provides details of our experimental setup, including HOCVP t...

    arXiv 2026

  79. [80]

    You will receive A set of images depicting specific humans and objects and a text description detailing the target patterns of human-object interaction

  80. [81]

    MLLM” indicates whether a multimodal large language model is incorporated into the training of video personalization to enhance generative performance. “Intra- subject Distinction

    Your task: analyze the precise relationships between the visual entities in the images and the interaction logic defined in the text prompt. Specifically, evaluate how the distinct human and object features from the visual inputs should be spatially, physically, and semantically integrated to accurately reconstruct the interaction described in the text. T...

    2048

Showing first 80 references.