MAVE turns film shots into editable 3D blockouts with a new benchmark
A research team proposes a video-to-blockout framework that reconstructs cameras, motion, and scenes into editable Blender files, plus a benchmark and training dataset.
In plain words
Filmmakers often build rough 3D stand-ins, called blockouts, to plan shots, but recreating them from finished video by hand is slow. MAVE is a system that watches a film shot and automatically rebuilds it as an editable 3D scene. The authors also released a test set and a training dataset to support this task.
What was announced
MAVE works in three stages, according to the paper's abstract. It first recovers camera motion, human movement, and geometric evidence within a shared metric coordinate system. An agent then builds an editable environment under constraints drawn from those observations, while programmatic checks flag inconsistencies to guide refinement. The result is an editable Blender scene and a frame-aligned blockout video, so creators can change viewpoints or character paths and render the edited scene to condition video generation.
For evaluation, the authors introduce MAVEBench, covering 50 film shots and 416 keyframes that test camera motion, spatial layout, and character blocking. They report that MAVE raises overall scores by 27.5 to 41.1 points across three agents compared with vanilla baselines, and with DeepSeek-1017 it reaches a score of 73.2 at an average API cost of USD 0.054 per shot while cutting total reconstruction time by 53 percent. The team also built MAVEData-3K, a set of 3,000 blockout-video pairs intended for model training, which they say will be released publicly; code is posted on GitHub.
What it does not show
- Evidence is limited to the paper's abstract, so the method, benchmark design, and reported scores cannot be verified in detail.
- All results, including cost and timing figures, are self-reported by the authors and have not been independently confirmed.
- The abstract does not state whether MAVEBench or the code have already been released, only that MAVEData-3K will be made public.
For security and IT teams
- Access
- The paper links to a public code repository on GitHub, and the authors state the MAVEData-3K dataset will be released publicly.
Sources
Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 19 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.