We wired ByteDance's Seedance 2.5 into our API platform. Integration notes and the gotchas nobody documents

Last week we got Seedance 2.5 running on the official Ark channel, and since Chinese docs on this model are thin and English docs are thinner, here are the notes I wish someone had written before we started.

The quick facts first. One async endpoint, model id `doubao-seedance-2.5-face`. Each request produces 4 to 30 seconds of video at 480p or 720p, MP4 or MOV out. Audio is generated in the same pass: speech, sound effects, background music. Spoken lines go inside double quotes in the prompt, everything else reads as scene direction. References go up to 30 images + 10 videos + 10 audio tracks per request, a lone audio file is a valid input, and real-person reference material is allowed, which is what the "face" in the id means.

Now the parts that cost us time:

- Prompt is mandatory on every mode, including image-to-video. Minimum 3 characters. 2.0 let you throw a bare image; 2.5 rejects it.
- The cheaper "reference" billing rate only triggers when you attach reference video. Image and audio references bill at the full text rate. Easy to misread, expensive to assume.
- With video input there is a minimum billing floor of roughly 5/3 of the output seconds, and input seconds count toward the bill. Trim your reference clips.
- The vendor supports duration=-1 (model picks the length), but there is no way to bill an unknown duration up front, so we could not expose it. Same reason the video-edit mode is not available in v1.
- No seed parameter. Two identical requests give two different videos.


Model page with playground and full parameter reference:
https://reapi.ai/models/seedance-2-5
https://reapi.ai

If you try it and hit an edge case I did not list, reply here. Genuinely curious what breaks first.