One 16:9 hero video cropped to 9:16 and 1:1 for $1.04

One 8-second 16:9 Wan 3.0 video at 720p ($1.00) plus two Video filter crops ($0.04) gives you a widescreen, a vertical and a square version for $1.04. A crop is a flat $0.02 per encode job, and the crop is described in fractions of the frame, not pixels.

Buyers who publish on several feeds often generate once and reframe. The reframe is a worker ffmpeg pass with no model inference, which is why it is cheap. Rates come from the Sume pricing code and the Video filter page (read 2026-10-10).

What is the endpoint chain?

Generate once, check each crop for free, then encode.

• POST /v1/videos with wan-3.0, duration: 8, resolution: 720p, aspect_ratio: 16:9. Keep the subject in the middle of the frame, because both crops come from the centre.
• POST /v1/video-filter/check with the crop program: this is free and returns valid plus diagnostics, with no job and no charge.
• POST /v1/video-filter with ops: [{op: "crop", x: 0.3418, y: 0, width: 0.3164, height: 1}] for 9:16. The video_url must be the generated media.sume.com URL.
• POST /v1/video-filter again with {op: "crop", x: 0.21875, y: 0, width: 0.5625, height: 1} for 1:1.

What do the crop numbers come from?

A crop rectangle is a share of the source frame. For a 16:9 source, the width of a 9:16 slice is (9/16) divided by (16/9) = 0.3164 of the frame width, centred at x = (1 - 0.3164) / 2 = 0.3418. A 1:1 slice is 0.5625 wide at x = 0.21875.

What are the limits?

Crop width and height must be between 0.05 and 1, and x + width and y + height cannot pass 1, or you get video_filter_crop_out_of_bounds. The compiler rounds to even values for yuv420p. A crop is not a scale: a 720p 16:9 source is 1280x720, so the 9:16 slice is about 405x720, which is soft for a full-screen phone feed.

For a sharp vertical, generate it as 9:16 directly: Wan 3.0 takes the aspect ratio, and the same 8 seconds costs the same $1.00. Use crop when one hero must serve several shapes and a centred subject is fine. The source must be 300 seconds or shorter and on media.sume.com. See the Video filter page for refusal codes.

When should I generate each version separately?

If the 9:16 version matters most, generate it natively and crop only the secondary shapes. The extra generation is $1.00 and a native shape avoids the 405-pixel width. For three shapes that is $3.00 against $1.04, so crop wins on price and loses on resolution; decide by where the video will be watched. Compare the 480p proof-then-final loop to test the framing cheaply first, and the first-frame still-to-clip post.

Pre-publish checklist

Run the free check call for each crop before you encode. It returns diagnostics instead of a 400 and creates no job, so a typo in a fraction costs nothing.

Look at each cut on a phone. If the subject is clipped, move x toward it; the crop does not follow the subject by itself.

• Keep the subject centred at generation time.
• Send a separate Idempotency-Key for every encode.
• Use the new video_url from each result, never the source.