OmniXRef: A Unified Framework for Multimodal Multi-Reference Video Editing

Xiaoyan Cong^1* Tianyu Wang² Hyeonho Jeong² Zhifei Zhang² Yuqian Zhou² Yijun Li² Shaoteng Liu² Zhe Lin² Srinath Sridhar¹ Soo Ye Kim^2†

¹Brown University²Adobe Research

^*Work done during an internship at Adobe Research^†Corresponding author

TL;DRWe introduce OmniXRef, a unified framework for multimodal multi-reference video editing.
Our omni-reference set-up supports four dimensions: (i) number of references (1, ..., K), (ii) reference modality (text, image, video), (iii) reference attribute (camera, motion, content, style, VFX) and (iv) composition (text-image, text-video, image-video, text-image-video).

For the best experience, watch with sound onBest with sound on

Explore examples View benchmark

DEMO

OmniXRef

Qualitative results

Gallery

Qualitative comparison

Comparison with Baselines

Quantitative results

Benchmark Leaderboard

MLLM (Gemini-3.8-Flash) evaluation on OmniXRef-Bench.RA = Reference Adherence, VQ = Visual Quality, and SP = Source Preservation.The best result in each column is highlighted.