arXiv 2606.23610 introduces Vera, a joint project between Netflix and Caltech that uses a "Mixture-of-Transformers" architecture to separate video editing into an "edit layer" and the "original video." The result: high-fidelity video editing with no degradation of the original content.
The technical details: Vera is a video editing model that takes two inputs — the original video and an "edit specification" (e.g., "change the sky to sunset," "add a person in the background"). It outputs a "edit layer" (a sparse set of edits) and the "original video," then composes them at render time. The "edit layer" is a low-rank delta to the original video, preserving the original quality.
The "Mixture-of-Transformers" architecture: a 3-expert MoE, where each expert specializes in a different type of edit (color, object, motion). The router dynamically selects the right expert per edit operation. This allows Vera to handle diverse edit types with a single model.
The benchmark: on the "video edit fidelity" benchmark (which measures how well the edit preserves the original content), Vera scores 92.3, on par with human editors. The "edit speed" is 5× faster than traditional NLE workflows (Premiere, DaVinci).
The bigger takeaway: "edit as a layer" is a powerful abstraction for video editing. Traditional NLEs edit the raw pixels, which always introduces quality loss. Vera's "edit layer + original" approach is non-destructive, and the result is indistinguishable from a human-edited video. For the industry, this means "AI video editing" is moving from "generate a new video" to "edit the existing video non-destructively" — a much higher-value use case.