DiasR: Dual-Modal Identity-Anchored Sparse Routing for Efficient Multi-Subject Video Generation

DiasR is an efficient framework for personalized multi-subject video generation, achieving state-of-the-art identity consistency with nearly constant inference time as the number of subjects grows.

Abstract

Personalized multi-subject video generation is a promising direction within the field of controllable video generation; however, existing methods face challenges in maintaining cross-frame identity consistency and incur high computational overhead. To address these issues, we propose DiasR, an efficient framework that integrates Dual-Modal Identity-Anchored Alignment and a novel Sparse Routing Strategy. The Dual-Modal Identity-Anchored Alignment employs learnable identity queries to align visual and textual modalities with ground-truth subject masks, thereby mitigating cross-frame identity drift. The Sparse Routing Strategy dynamically routes video tokens to relevant subjects and groups them through bucket aggregation, reducing computational overhead and alleviating identity entanglement induced by redundant tokens. We also construct MuSA-2M, a large-scale dataset comprising 2 million annotated samples equipped with subject-level masks, which fills the gap in existing multi-subject video datasets. Experiments conducted on the OpenS2VEval benchmark demonstrate that our method achieves superior performance in identity consistency, text fidelity, and video naturalness. Notably, it maintains a nearly constant inference time as the number of reference subjects increases, outperforming existing baselines in both efficiency and generation quality for scenarios involving multi-subject interactions.

Method

Framework

Framework of our DiasR. It takes noisy latents, multi-subject reference images, and text prompts as inputs, and integrates a Dual-Modal Identity-Anchored Alignment mechanism to anchor the identity of each subject, together with a Sparse Routing Strategy that mitigates the extra computational overhead induced by subject tokens.

Sparse Routing Strategy

Left: Subject-level routing selects the most relevant subject for each query token. Right: Bucketed attention computes attention independently for each subject bucket, avoiding performance degradation while preserving cross-subject interactions.

Paired Concept Disentangled Dataset

Dataset construction pipeline of MuSA-2M.

Experiments

Qualitative Comparison

DiasR consistently preserve the identities of multiple subjects and correctly models interactions across different subject groups.

Quantitative Comparison

On OpenS2V-Eval, DiasR achieves the best performance on FaceSim (60.66%) and NexusScore (47.40%), and reaches the highest overall TotalScore (57.79%), demonstrating superior identity preservation and generation quality.

Computational Performance Comparison

The sampling time and the floating-point operations (FLOPs) versus the number of reference subjects, DiasR exhibit nearly constant inference time as the number of reference entities increases.

Human Preference Study

We conduct a human preference study on Identity Consistency, Video Quality, Text Fidelity, and Interaction Naturalness. DiasR ranks first in Identity Consistency and Interaction Naturalness

BibTeX

@inproceedings{li2026diasr,
    title={DiasR: Dual-Modal Identity-Anchored Sparse Routing for Efficient Multi-Subject Video Generation},
    author={Li, Yangyang and Liu, Wu and Li, Jie and Liu, Xinchen and Zhang, Yongdong and Jin, Guoqing},
    booktitle={International Conference on Machine Learning},
    year={2026}
}