ASID-Caption

We build ASID-Caption, a data-and-model suite for fine-grained audiovisual video understanding.

Our goal is to move beyond “one video → one generic caption” by providing attribute-structured supervision and quality-verified annotations, enabling models to produce more complete, more controllable, and more temporally consistent descriptions that cover both visual content and audio cues.

What we release

• ASID-1M: a large-scale collection of attribute-structured audiovisual instructions with both single-attribute and all-attributes training formats.
• ASID-Verify: a scalable curation pipeline that generates, ensembles, verifies, and refines annotations to improve semantic and temporal consistency.
• ASID-Captioner: Qwen2.5-Omni-based audiovisual captioning models fine-tuned on ASID-1M.

Research interests

• Video understanding & video captioning
• Audio-visual learning
• Multimodal LLMs / instruction tuning
• Data curation, verification, and quality control