Research Tests the Limits of AI-Assisted Video Editing and Generation
RoughCut-Bench evaluates vision-language models on rough-cut assembly and transcript cleanup, with released code and prompts but reported limitations. A separate direct-pixel-prediction paper reports a 23% reduction in warping error on Wan; independent replication and practical creator benefits remain unproven.
Sources (3)
Updated Nov 19, 2026