AI video editing can now fake geometry and light — at least according to researchers behind InsertAnywhere.
The framework tackles video object insertion, a task that sounds simple but has stumped production pipelines for years. The system lets a user anchor an object in a single frame, then automatically tracks its placement across the rest of the clip, handling occlusions and local scene motion without manual correction. The harder problem — making inserted objects interact with light realistically — gets addressed through a technique the authors call Optics-Aware Representation Alignment, which extends feature extraction beyond the object's edge so shadows and reflections bleed into surrounding pixels the way they would in a real scene. To train that capability, the team built and open-sourced ROSE++, a dataset of quadruplet samples specifically designed for supervised learning of optical effects.
The gap between "generated video" and "composited video" has been a quiet embarrassment for AI editing tools: objects that float unnaturally or cast no shadow break the illusion immediately. If InsertAnywhere's claims hold up outside controlled benchmarks, it would close that gap for effects that currently require dedicated VFX software and artists.
The paper reports that InsertAnywhere outperforms both existing research methods and commercial generative tools in geometric plausibility and photometric realism — though "significantly outperforming" in a lab setting and actually replacing a VFX pipeline in production are two different things.