Personalized video generation faces a significant scalability challenge: as the number of custom subjects increases, existing methods suffer from severe identity leakage and semantic entanglement, resulting in videos with blended attributes and incoherent narratives. We identify that this stems from a lack of explicit binding between textual entity phrases and their corresponding visual patterns prior to fusion in diffusion models. To address this, we propose MultiBinder, a novel framework built on a bind-then-narrate'' paradigm. Our approach introduces a two-stage conditioning process before video diffusion. First, Cross-modal Entity Binding devises an entity-aware Q-Former to extract identity-centric features from each subject's reference image and tightly align them with its unique entity phrase, yielding disentangled identity tokens.'' Second, a Multimodal Large Language Model (MLLM) acts as a narrative director'' in the Narrative-Aware Subject Arrangement stage, capturing the action of each subject from the text prompt and synthesizes contextualized subject features that encapsulate who is doing what'' and preserve role-specific distinctions. These features are then injected into a video diffusion model via dedicated conditioning pathways, ensuring precise attribute allocation. Extensive experiments on both general and newly curated multi-subject benchmarks show that MultiBinder significantly outperforms existing methods in multi-subject identity fidelity.
Figures & tables
Figure 1. Given a text prompt and four subjects, conventional approaches (top) suffer from ambiguous cross-modal alignment, leading to identity entanglement where visual attributes blend incoherently across subjects (e.g., the “elderly man” incorrectly inherits the face of ID1). MultiBinder (bottom) establishes unambiguous entity bindings, ensuring strict identity preservation.
Figure 2. Overview of the MultiBinder framework. The Cross-modal Entity Binding module utilizes an Entity-aware Q-Former and Entity-Refinement Attention to link visual attributes with entity phrases. Subsequently, the MLLM performs Narrative-Aware Subject Arrangement, interpreting the text prompt to contextualize identity tokens for the final Diffusion Transformer.
Method
Aes. ↑
M. Smo. ↑
M. Amp. ↑
FaceSim ↑
Gme. ↑
Nexus. ↑
Natural. ↑
Total. ↑
VACE-P1.3B
47.34%
96.8%
12.03%
16.59%
71.38%
40.19%
64.31%
48.98%
VACE-1.3B
48.24%
97.2%
18.83%
20.57%
71.26%
37.91%
65.46%
49.89%
Phantom-1.3B
46.67%
93.3%
14.29%
48.56%
69.43%
42.48%
62.5%
54.89%
MultiBinder
48.13%
97.06%
32.16%
43.45%
69.89%
43.17%
65.18%
55.52%
Table 1. Performance comparisons with state-of-the-art approaches on the OpenS2V-Eval open domain in terms of Aesthetics, Motion Smoothness, Motion Amplitude, Face Similarity, GmeScore, NexusScore, and NaturalScore. The best-performing result is highlighted in bold , and the second-best is underlined .
Method
SIPS↑
SVRS↑
VACE-P1.3B
46.37%
61.20%
VACE-1.3B
54.64%
63.09%
Phantom-1.3B
55.03%
72.15%
VACE-14B
62.38%
77.10%
Phantom-14B
65.46%
76.44%
Kling-1.6
69.51%
82.89%
Table 2. Performance comparisons with state-of-the-art approaches including open-source baselines with varying parameter scales and the commercial system Kling-1.6 on MultiBind-Bench. Results are reported in terms of Identity Purity Score ( SIPS ) and Visual Reasoning Score ( SVRS ). The best-performing result is highlighted in bold , and the second-best is underlined .
Figure 3. Examples of videos generated by different approaches with multi-subject reference images.
Figure 4. Examples of videos generated by different approaches on scenarios with an increased number of subjects.
Method
TotalScore ↑
SIPS↑
SVRS↑
Cross-Modal Entity Binding
w/o Q-Former
49.21%
51.63%
63.11%
Vanilla Q-Former
49.95%
52.75%
64.20%
Entity-aware Q-Former
52.43%
56.31%
73.69%
Entity-aware Q-Former + ERA
53.17%
59.46%
73.85%
Narrative-Aware Subject Arrangement
Table 3. Ablation studies on the impact of each design in MultiBinder. Training in this table uses 50% data.
Figure 5. User study between MultiBinder and other two state-of-the-art models across three dimensions. MultiBinder exhibits competent performance in subject consistency and text alignment.
Recent advances in diffusion models have significantly improved text-to-video generation, enabling personalized content creation with fine-grained control over both foreground and background elements. However, precise face-attribute alignment across subjects remains challenging, as existing methods lack explicit mechanisms to ensure intra-group consistency. Addressing this gap requires both explicit modeling strategies and face-attribute-aware data resources. We therefore propose LumosX, a framework that advances both data and model design. On the data side, a tailored collection pipeline orchestrates captions and visual cues from independent videos, while multimodal large language models (MLLMs) infer and assign subject-specific dependencies. These extracted relational priors impose a finer-grained structure that amplifies the expressive control of personalized video generation and enables the construction of a comprehensive benchmark. On the modeling side, Relational Self-Attention and Relational Cross-Attention intertwine position-aware embeddings with refined attention dynamics to inscribe explicit subject-attribute dependencies, enforcing disciplined intra-group cohesion and amplifying the separation between distinct subject clusters. Comprehensive evaluations on our benchmark demonstrate that LumosX achieves state-of-the-art performance in fine-grained, identity-consistent, and semantically aligned personalized multi-subject video generation. Code and models are available at https://jiazheng-xing.github.io/lumosx-home/.
Jiazheng Xing, Fei Du, Hangjie Yuan +7
Zhejiang University · National University of Singapore · DAMO Academy, Alibaba Group +1
Subject-driven and multi-element video generation are central to controllable video synthesis, but existing methods still struggle to preserve identity consistency and model complex relationships among multiple subjects. In this paper, we propose Aura, a unified framework for high-fidelity and identity-consistent video generation. To better capture scene dynamics and subject interactions, we introduce AI director-level captions that provide dense and structured descriptions of video content. We further leverage a vision-language model (VLM) with learnable queries to extract multimodal semantic features from textual and visual references, covering both global semantics and fine-grained visual cues. To bridge the representational gap between the VLM and the Diffusion Transformer (DiT), we design a two-stage alignment strategy that progressively maps VLM features into the DiT feature space. For visual conditioning, we adopt token concatenation to inject reference information directly into the generation process. To distinguish heterogeneous subject types and reduce common copy-paste artifacts, we develop a subject-aware RoPE-Shift mechanism. To further differentiate reference images of different categories, we introduce subject-aware learnable tokens. In addition, we introduce Memory Tokens to balance the training signal across examples with different numbers of reference subjects. During inference, Progressive-APG (Adaptive Prompt Guidance) further alleviates oversaturation and improves semantic alignment with user prompts. Finally, we build a high-quality video-subject image dataset through a dedicated data construction pipeline. Extensive experiments show that our method achieves state-of-the-art performance on both single-subject generation and more challenging multi-element scenarios.
Recent advances in video synthesis have enabled realistic integration of real individuals, driving demand for identity-aware generation. While emerging methods support joint appearance and voice injection in audio-visual models, they primarily focus on single-subject settings. Multimodal identity integration across multiple subjects remains limited, and precise alignment between visual and vocal identities in multi-subject scenarios remains underexplored. We present Identity-as-Presence, a unified framework for joint personalized audio-video generation. An automated data curation pipeline constructs identity-labeled audio-visual pairs for single- and multi-subject scenes. A unified identity injection mechanism then binds paired appearance and voice through shared cross-modal identity binding and subject-anchored captions. A multi-stage training strategy further leverages large-scale unimodal data alongside scarce paired clips to mitigate modality imbalance. Experiments show superior audio quality, video fidelity, and audio-visual consistency, with stronger multi-subject binding than the compared methods. For more details and qualitative results, please refer to our webpage: Identity-as-Presence.