A new technique called CleanVideo can strip banned or unwanted content out of AI-generated video without leaving the rest of the clip warped or jittery.
The method, described in a new arXiv paper, targets text-to-video diffusion models, the systems behind tools that turn a written prompt into a moving clip. Instead of applying one blanket fix, CleanVideo uses a tri-modal gating system that reads the video's spatial and temporal features, the diffusion model's timestep, and the text prompt itself to decide exactly where, when, and whether to intervene. When a clean substitute concept exists, it steers the banned content toward that substitute rather than just deleting it outright. Tested on three separate video diffusion models, the approach beat existing erasure baselines on both frame-by-frame and whole-clip evaluations, and it held up even when attackers tried to coax the erased concept back out.
Video is a harder erasure problem than static images because a concept can drift in and out of frame, change appearance across the clip, and shift with each denoising step, so a fix tuned for one frame can fail on the next. That gap is why concept removal in image generators has moved faster than in video models, and why platforms shipping video generation to the public need something more precise than a single global filter.
Earlier attempts at scrubbing concepts from video often left behind a visible blur or stutter every time the model tried to censor itself, trading one flaw for another. CleanVideo's targeted approach suggests that tradeoff is not inevitable, though the real test will be whether it holds up against concepts nobody thought to name in the paper.