CutCraft benchmark: 13 video models fail pro editing tests

Modern multi-shot video generators can produce coherent output without being able to execute professional editing techniques, according to a new benchmark paper on arXiv introducing CutCraft. The team, led by Tianyi Zeng and colleagues, tested 13 state-of-the-art models, both closed and open-source, on shot structure, transition grammar, audio-video synchronization, and montage composition.

"Coherence does not imply the ability to execute editing techniques," the authors write. Across the tested systems they report a "consistent gap between coherence and editing-technique execution," alongside "unstable shot structures, weak transition control, and poor montage performance."

CutCraft combines four evaluation axes: "shot-structure alignment, expert-model metrics, tool-grounded multimodal judgment, and rubric-based question answering." It targets specific professional cuts including J-cuts, L-cuts, and transition timing. Aesthetic quality, the paper notes, correlates only weakly with editing compliance. A shot that looks polished can still violate the grammar an editor would apply.

The abstract does not publish a per-model leaderboard or per-technique numbers, and does not name which of the 13 systems score at the top or bottom. Two researchers in our tracked expert community shared the paper's link the day it went up.