Text description of the desired background after object removal.

Text description of what should be masked in the input video, such as the object or person to remove. Used to generate a temporary mask video with SAM-3 when `quad_mask_video_url` is not provided.

Negative prompt to guide generation away from undesired outputs.

URL of the input video containing the object to remove.

URL of a mask video for the removal target. For best results this should be a VOID-style quadmask video with 4 grayscale values: 0=object to remove, 63=overlap, 127=affected region, 255=background to keep. A simple binary mask (0=remove, 255=keep) also works. If omitted, the app will generate a temporary mask video from `mask_prompt` using the SAM-3 video endpoint and convert it to a quadmask automatically.

Run VOID Pass 2 warped-noise refinement after Pass 1. This is slower but can improve temporal consistency on longer clips.

Classifier-free guidance scale.

Temporal window size for inference. The backend snaps this to the nearest CogVideoX-safe value that works with temporal compression and patching. Valid outputs are 69, 77, 85, ..., 197.

Number of denoising steps. Higher values improve quality but increase latency.

Random seed for reproducibility.

Denoising strength. 1.0 means full denoising.