Skip to main content
Back to timeline
arXivSource publication:

Multi-subject video generation stops blending faces: MultiBinder binds each subject to its phrase before the narrative takes over

Synopsis

Personalized video generation degrades as the number of custom subjects grows: visual attributes of different people blend together and the intended narrative collapses. MultiBinder splits conditioning into two stages, first binding each reference image to its entity phrase with an entity-aware Q-Former to produce disentangled identity tokens, then letting a multimodal large language model infer who is doing what from the prompt before injecting the result into a video diffusion model. On the new MultiBind-Bench, MultiBinder scores 73.98% on Identity Purity and 84.56% on Visual Reasoning, above Kling-1.6 at 69.51% and 82.89% on the same benchmark.

Source-provided article image: MultiBinder: Preventing Multi-Subject Identity Leakage in Personalized Video Generation
arXiv

Interpretation

Multi-subject personalized video generation can now bind each subject's appearance to its own textual phrase before diffusion, reducing the blending of features across characters. Earlier methods fused reference images and the text prompt into a joint conditioning signal without isolating, before fusion, the correspondence between a given entity phrase and its visual pattern. On the newly built MultiBind-Bench, MultiBinder reaches 73.98% Identity Purity and 84.56% Visual Reasoning, above the commercial system Kling-1.6 at 69.51% and 82.89% on the same benchmark.

The pipeline runs in two steps, binding identities and then arranging the narrative, both before video diffusion. The first step uses an entity-aware Q-Former, a cross-modal encoder that pulls image features through text queries, to extract identity-centric visual features from each reference image and realign them to the phrase with Entity-Refinement Attention, yielding identity tokens. The second step has a frozen multimodal large language model act as narrative director, reading the prompt and reference images to produce contextualized who-is-doing-what features, which pass through connectors into a diffusion transformer built on the Wan2.1-T2V-1.3B backbone.

Ablations attribute the gains to specific modules rather than to using a large model in general. Replacing the binding module with a vanilla Q-Former gains only 0.74% in total score over the no-Q-Former baseline, while the entity-aware Q-Former gains 3.22%. On a 50% subset of the training data, adding Entity-Refinement Attention raises Identity Purity by 3.15%; narrative arrangement over a direct-condition baseline without the large model raises total score by 2.10%, Identity Purity by 4.91% and Visual Reasoning by 3.71%.

Perspective

Evaluation covers 180 subject-text pairs in OpenS2V-Eval and 100 multi-subject entries with 120 reference images in the self-built MultiBind-Bench, each entry holding 2 to 4 subjects; a qualitative example adds a five-subject scene. The method uses Wan2.1-T2V-1.3B as the video backbone and Qwen2.5-VL-3B-Instruct as the narrative model, trained on 500k videos with multi-object annotations.

The authors attribute identity leakage to the absence of explicit phrase-to-visual binding before fusion in diffusion models, a causal explanation supported by ablations rather than measured directly. MultiBind-Bench's Identity Purity relies on Qwen2.5-VL multiple-choice judgments over hair, face and clothing attributes, and Visual Reasoning is scored by GPT-4o on a 5-point scale averaged over two independent chain-of-thought runs, so both metrics rest on model judges.