remove_background is a video key, not a photo cutout

remove_background is a video key. It is not a photo cutout. The listed models on this tool are video background-removal models — pass the clip as video_url. For a still, remove the background from a photo routes through image-edit. Send a PNG into the video row and you have priced a clip model to stare at one frame. Send a talking-head MP4 into a still editor and you have one transparent portrait and a video you still cannot composite.

Remove the background from a video is the named agent job. Two paths, depending on what the footage looks like. Solid black or near-black (typical of AI talking heads) can take a fast FFmpeg color-key — remove_video_background — a flat processing fee. Anything else picks a dedicated video removal model, priced per model. Tell the agent the backdrop color so it picks the right path. The FFmpeg path will not save a green screen, a white cyclorama, or a real room.

What the job is

A clip whose subject you want isolated or composited. Output is the video with background removed or replaced. Then you overlay it, stitch it, or drop it on a new plate — those are later jobs.

Ask first whether the backdrop is black. Confirm. The supporting tool is explicit: it does not work for green, blue, white, or photographic scenes, and it should not be used as a silent default.

Cost: FFmpeg key is processing; AI removal is a model confirm. That split is the product. Ignoring it is how a black talking-head pays an AI segmenter it did not need, or a real-scene clip pays FFmpeg for a key that cannot exist.

What it is not

• Photo cutout. Image-edit models via generate_image_from_image. Different capability, different file type, different price.
• A generate of a subject on transparency. Prompting text-to-video for "alpha background" is not a key. You will get a gray void and still need this job, or you will not get a void and still need this job.
• The overlay itself. Overlay a talking head composites. It can remove a black overlay background as a flag. It is not a general keyer. Key first when the backdrop is not that special case.
• Green-screen fantasy on location footage. The AI video-removal path is for "other than pure black," not a promise that a busy street becomes a studio. If the model cannot hold the hair, do not loop the same row hoping. Change the plate.

Using the photo job on a video, or the video job on a photo, or a generate instead of a key, all waste a confirm. The composite you wanted is still unmade.

The test

Is the file an MP4 (or other video URL), and is the only remaining work isolation of a subject that is already in frame?

Yes: this job. Say whether the backdrop is black. Confirm the path.

If the file is a photo, leave. If the subject does not exist yet, generate a talking head or a scene first, then key. Order is the credit discipline.

FAQ

My talking head is on black. Do I still need an AI model?

Usually no. That is the FFmpeg path. Confirm it is actually black, not dark gray with a gradient. Near-black works; a bedroom at night does not.

Can I key a still by uploading it as a one-frame video?

You can torture the pipeline. You will pay a video path for an image job. Use the photo capability.

Does this replace shooting on a real green screen?

It replaces a black matte cheaply, and it attempts other backdrops via a model. It is not a promise about spill, hair, or handheld location. Plan the plate as if the key can fail.

Why not generate the person already on the new background?

Then identity, lighting, and the new plate are one roll of the dice. Keying a locked take lets you change the plate without rebuying the performance.