Cut out a subject from a video frame: video frames, then RMBG

To cut a subject out of a video, take one frame with Sume video frames as a lossless PNG, then send that frame to RMBG 1.0, which returns a PNG with alpha. Frame extraction is unbilled and the removal is one billed generation. This produces a still cutout, not a transparent video: Sume has no whole-video background removal.

Facts below are from Video frames, the RMBG route description, and MCP tools and gates.

Why a PNG frame?

Video frames returns jpeg by default or png for lossless inspection. A cutout depends on clean edges around hair and fabric, and JPEG block noise gives the model false edges to follow. Ask for format: "png" and leave max_edge out to keep the source frame size. Frames are durable media.sume.com image artifacts, which are HTTPS URLs RMBG accepts as image_url.

# 1) one lossless frame at 8.4 s, source size
curl -X POST https://api.sume.com/v1/video-frames \
  -H "Authorization: Bearer $SUME_API_KEY" -H "Content-Type: application/json" \
  -H "Idempotency-Key: frame-001" \
  -d '{"video_url":"https://media.sume.com/artifacts/artf_demo/talk.mp4","at":[8.4],"format":"png"}'

# 2) poll GET /v1/video-frames/$REQUEST_ID, then cut out the frame URL
curl -X POST https://api.sume.com/v1/rmbg-1.0/remove \
  -H "Authorization: Bearer $SUME_API_KEY" -H "Content-Type: application/json" \
  -H "Idempotency-Key: cutout-001" \
  -d '{"image_url":"https://media.sume.com/artifacts/artf_frame/frame-8.4.png"}'

How do I pick a frame worth cutting?

Request several times at once: at takes 1 to 24 values, and each must be at least 0 and below the duration. Look for a frame where the subject is sharp and the hands or hair are not blurred by motion. Spending unbilled frames on a shortlist is cheaper than paying for removals on bad ones.

The source clip has to be 300 seconds or shorter for video frames. A frame whose extract failed returns a null url and the job still succeeds, so check each entry before you send it on.

Where does the cutout go?

A cutout is a common thumbnail ingredient. YouTube's thumbnail page, read 2026-10-02, recommends 2160 x 3840 for Shorts and 3840 x 2160 for videos, with JPG or PNG as formats. A transparent PNG is a layer you place on a background in your design tool; Sume's RMBG output is the layer, not the composed thumbnail.

If the frame already has transparency from a previous cut, RMBG flattens it onto mid-gray #808080 before sending it to the provider so edges stay clean. That is still one billed removal.

What does it cost and how long does it take?

Frame extraction is unbilled and runs as an async job: submit returns 202, and you poll GET /v1/video-frames/:id or the generic job status route. RMBG is one billed generation per image; the exact public rate is in GET /v1/catalog. A shortlist of eight frames and two removals costs two removals, which is the reason to shortlist.

Both calls need an Idempotency-Key, and you should use a different one for each different request. On hosted MCP the same flow is video_frames_create, jobs_wait, video_frames_get, then rmbg_create, jobs_wait, jobs_result, with idempotency_key on the writes.

After the cutout, inspect the edges at 100 percent on both a light and a dark background. Hair against a bright window is the usual failure. If the cutout is poor, the cheapest fix is a different frame, not a different tool: step 0.2 seconds either side, where the motion blur and lighting differ, and try again.

What Sume does not do

Check pricing for the removal in GET /v1/catalog before you run a batch, since it is billed per removal.

• No background removal across a whole video. For a green screen, use the video filter chromakey path instead.
• No matting control: RMBG takes an image URL and returns a PNG, with no parameters for edge softness.
• No composition: placing the cutout on a new background is a separate step.