Test five caption styles on one clip for $1.50 on Sume

Testing five caption styles on one finished clip costs $1.50 on Sume: one 5-second MiniMax H3 Max clip at $0.50 plus five caption jobs at $0.20. Regenerating the clip for each variant would cost $3.50. Because captions are burned onto an existing video and priced flat, the clip is generated once.

Why the numbers differ

Video captions are a flat $0.20 per job for videos up to 60 seconds, whatever the clip's own cost. A caption job takes an existing video and returns a captioned one, so you can run it repeatedly on the same source. Regenerating the clip for each style would also change the footage, so you would be testing the clip and the caption together and learning nothing clean.

What to vary

Change one thing at a time. The captions docs name the style values slam, punch and tiktok-green, and a design field overrides a style's colours, typography, placement, phrasing and motion for one request. Three named styles plus two design overrides of slam make five variants; design is not supported on punch or tiktok-green. See the captions page for the option details.

For a silent clip, pass cues with your own text. A silent clip without cues fails with caption_no_speech, which wastes no money but wastes time.

When five is too many

Five variants at $1.00 total is cheap next to the clip, but the real cost is review time. Start with two variants, pick a winner, and test the next variable on the winner. Two rounds of two cost $0.80 and give you more information than one round of five. If the clip is a Seedance 2.5 hero at $2.889, the same logic applies: the caption cost stays at $0.20, so the caption test is a smaller share of the bill.

A retry reserve is not needed for captions in the same way as for generation, because the clip is already paid for. Keep $1 aside and run the test.

• Generate the clip once and keep its URL.
• Run variants as separate caption jobs against that source.
• Name outputs by the variable changed.
• Pick a winner before testing the next variable.

A budget for the whole experiment

Add a second clip to learn whether the winner holds on different footage. Two clips with three styles each cost $1.00 for the clips plus $1.20 for six caption jobs, $2.20 in all. If the same style wins on both, you have a default for the series. If not, the footage matters more than the style, and you have learned that for $2.20.

Keep the source video URL public and stable while you run the variants, since captions are burned onto an existing public video URL.

Reading the results

Judge each variant on the device your viewers use, with sound off if your feed is mostly muted. Check three things: can you read every line at arm's length, does the text cover the subject's face or the product, and does the timing keep up with the cut. A style that wins on a desktop preview can lose on a phone.

Keep the loser files. A style that fails on one clip is sometimes right for another format, and you have already paid $0.20 for it.