netflix / void-model
void-model
Netflix VOID Model removes objects from videos and rewrites shadows, reflections, and motion after the edit. Use it for VFX cleanup and counterfactual scene edits.
## Overview VOID Model is a video-to-video editing model released by Netflix. It removes a target object and also rewrites the interactions it caused. That includes shadows, reflections, and downstream motion like collisions or falling objects. It uses a video diffusion backbone (CogVideoX-Fun) with an interaction-aware mask signal, which helps it keep motion and scene dynamics coherent. ## What you can build - Remove a person from a shot while keeping the rest of the scene believable - Clean up props or set mistakes without breaking nearby motion - Create “what if it wasn’t there” counterfactual clips for editing and previs - Delete an object that was blocking or pushing something, then regenerate the aftermath - Produce VFX plates where secondary effects still match the new scene ## Inputs - A source video file in MP4 format. Short, steady clips work best. - An edit description that states what the scene should look like after the change. Describe what remains in the frame. - An optional exclude list that names visual traits you don’t want in the result. - A quality-versus-speed control that sets how many denoising iterations the model runs (allowed range: 1 to 50). - An optional random seed for repeatable results (allowed range: 0 to 9,999,999). ## Outputs - One edited MP4 video that applies the requested change to the input clip. - The output preserves the idea of the original shot while regenerating the altered regions. - When the edit implies physical consequences, the output can change object motion to stay plausible. ## Limitations - Video diffusion models can show “object morphing,” where shapes wobble across frames. VOID includes a second-pass refinement in its reference pipeline, but artifacts can still appear. - Results depend on accurately identifying what should change versus what must stay. If the target region is ambiguous, the edit can spill. - Heavy motion blur, fast camera moves, and strong occlusions reduce temporal stability. - Low-quality inputs increase risk of flicker and texture drift. This includes compressed MP4s, noisy footage, and dark scenes. ## Safety & compliance - Only edit videos you own or have rights to modify. - Don’t use it for deception, impersonation, or to mislead viewers about real events. - Get consent before removing people or altering identifying features in private footage. - Don’t use it to remove watermarks, logos, or safety labels. - Disclose that a clip was edited when accuracy and trust matter.
API quick start
Run void-model with a single API call.
{
"prompt": "A lime falls on the table.",
"inputVideo": "https://your-cdn.com/input.mp4",
"removeObjects": "glass",
"affectedObjects": "lime"
}