Research framework for real-time, causal streaming video editing with diffusion, running at about 12.66 FPS.
Updated EU status checked Last shown
Sources: the vendor's own pages (linked above) and creator videos at the linked timestamps. EU status is information, not legal advice; see how we check.
LiveEdit is a research framework for diffusion-based streaming video editing, presented as an ECCV 2026 paper by Xinyu Wang, Chongbo Zhao, Fangneng Zhan and Yue Ma. It edits video causally, frame by frame (in chunks), following a text instruction, while keeping the background and non-edited regions stable over time.
The method trains a bidirectional diffusion transformer with full temporal attention as an editing teacher. A three-stage distillation pipeline then converts it into a chunk-wise causal streaming editor, using teacher forcing, and compresses generation to 4 diffusion steps. An autoregressive-oriented mask cache reuses computation for masked background tokens across frames. The authors report state-of-the-art quality among streaming baselines and 12.66 FPS inference. A YouTube creator described the speed as nearly 13 fps and said code and a training script were released, about 17 GB in total.
It is aimed at researchers and developers interested in interactive or live video editing. The provided pages do not give a repository URL, model weights location, license or hosted demo, so those details are unconfirmed. Limits are those of a research release.
Watch the moment: the exact second top creators show it. Videos load only when you press play, via youtube-nocookie.com.
Edits videos in real time at nearly 13 fps using a causal chunk-wise streaming setup, driven by a prompt. Code and training script are released, about 17 GB total.
Closest neighbours by meaning, not just by category.
Model · JD
Open-weights 16B autoregressive diffusion model from JD for real-time, instruction-guided editing of streaming video.
Research · ByteDance
ByteDance research method for post-training image diffusion models using on-policy self-distillation from reward signals.
Model
Research model for video relighting that jointly generates relit video and per-frame environment maps, no camera poses needed.
Model
Research project for real-time autoregressive generation of multi-shot video narratives from shot-level prompts.
Model
Research method for multi-shot text-to-video with frame-accurate cuts, fades and camera moves, built on Wan2.1-T2V-14B.
Library
Training-free staged sampling method that speeds up pretrained flow-matching text-to-image diffusion models.