Peking U & Kling Team Release RefCaptioner, Tackling Video Understanding Hallucinations

jiqizhixin · x · 2026-08-19

Video models often hallucinate, e.g., confusing a woman's red dress with a distractor dog. Researchers from Peking University and the Kling Team introduce RefCaptioner, designed to ground visual phrases to specific reference images accurately. Their two-stage recipe combines supervised fine-tuning on mixed data with a custom RL step (HCD-GRPO) that rewards correct phrase-image binding and penalizes hallucinations. Trained on 20K videos and 171K reference images, the model achieves SOTA on the new MRVBench, outperforms open-source rivals, and produces human-preferred captions along with more faithful video generation.

Original post →

More from Multimodal

Multimodal channel →