Omnishow

OmniShow is an open-source end-to-end human-object interaction video generation model developed by ByteDance together with multiple universities. It supports unified multimodal input of reference image, audio, posture and text. With innovative attention mechanism and training strategies, it optimizes interaction details such as grabbing and holding objects, avoiding motion distortion and object penetration, and maintains high consistency of human appearance and costumes while supporting long-duration fluent video generation. It reaches industry-leading performance on professional benchmarks, suitable for e-commerce virtual live streaming, AI short dramas, brand digital human promotion, virtual idol performance and game character animation. The open-source model allows local deployment and secondary development.