UniVid Open-Source Model Unifies Video Generation and Understanding, Gains 2.2% on VBench-Long
- UniVid couples an MLLM to a diffusion decoder through a lightweight adapter, so one model does both video understanding and video generation.
- Temperature Modality Alignment corrects text-token imbalance in flow-based generation and improves how closely output follows the prompt.
- Pyramid Reflection picks keyframes dynamically for temporal reasoning instead of processing every frame, cutting the cost of video question answering.
- UniVid claims state-of-the-art results: a 2.2% higher VBench-Long total than EasyAnimateV5.1, plus 1.0% on MSVD-QA and 3.3% on ActivityNet-QA over the best prior 7B baselines.
- The work is arXiv:2509.24200 by Jiabin Luo and 6 coauthors at UTS and other institutions, with the paper licensed CC BY-NC-SA 4.0 and a v2 filed the next day.