ICML 2026｜Meitu proposes BridgeRemoval to make video elimination no longer confusing

Intelligent Stuff Author | ZeR0 Editor | Mo Ying Eliminating passers-by and clutter in photos is now a "must-have artifact" for holiday productions. The ability of AI elimination has gradually expanded from pictures to videos. However, it must not only complete the content of a single frame, but also ensure the spatial coherence within a single frame and the timing consistency between frames, making video target removal significantly more difficult. The current mainstream method is based on the generation paradigm of "from noise to data" based on the diffusion model, that is, using uninformative Gaussian noise as the starting point for gradual reconstruction. This process discards a large number of effective structure and scene priors in the source video. When faced with large-scale occlusion or the need to generate content with complex physical logic, it often results in incomplete elimination of objects or the generation of content that is inconsistent with the scene. Therefore, it is generally solved by introducing a reference frame guidance mechanism, but this mechanism significantly increases the complexity of the system. In this regard, Meitu Imaging Research Institute (MT Lab) proposed a new framework based on the random bridge model - BridgeRemoval, which defined "video target removal" as a "video-to-video translation task" for the first time. By constructing a direct generation path based on the random bridge (VP-SDE Bridge), it can better balance background fidelity and generation flexibility, and significantly improve the accuracy of video target removal and the spatio-temporal logical consistency of the picture content. This result has recently been accepted by ICML 2026 (International Conference on Machine Learning), one of the three top conferences in the field of machine learning. Paper link: https://arxiv.org/pdf/2601.12066 Project homepage: https://bridgeremoval.GitHub.io/ The architecture of BridgeRemoval is to project the source video input into the hidden space through the frozen VAE encoder, based on VP-SDE Bridge Construct an interpolation trajectory directly from the source video prior (z_src) to the clean target (z_tgt), and then splice the mask latent variable (z_M) with the source video latent variable (z_src