One continuous take with synchronised audio, rendered in a single pass — no stitching, no scene-cut splicing. The barista locks the portafilter, pulls the shot, lifts the cup and turns to camera, all in one unbroken shot.
Seedance 2.5 seedance-25-t2v — 8s at 720p, one pass, rendered in 203.6s. No concat step exists in this pipeline because there is nothing to join.
ffmpeg-concat. That is cheaper and usually right — but the joins drift in colour and subject, and the audio has to be laid over the top because no clip shares a bed with its neighbour. Seedance 2.5 renders the whole span at once, so continuity is not something you repair in post; it is a property of the render.Baselined on the production MCP before shipping — the numbers the routing and the timeouts are built from.
| Probe | Result | Wall clock |
|---|---|---|
| t2v · 4s · 480p · n=5 | 5/5 | p50 187s · p95 193s |
| t2v · 12s · 480p | 1/1 | 230s |
| t2v · 8s · 720p (the clip above) | 1/1 | 204s |
| ref2v · 4s · 480p · n=3 | 3/3 | p50 194s · p95 196s |
| i2v · 4s · 480p · n=3 | 2/3 | p50 188s (one abort, see below) |
Latency is not linear in duration: about 165s of fixed overhead plus ~5.4s per second of video, which holds across the resolution change. So a 30-second take lands near 330s rather than the ~1,400s a linear reading would predict — and that single fact is what makes a 30s take practical here at all.
The one i2v abort was our ceiling, not the model: a slow tail ran past the 250s default timeout and was cut off, billed, and discarded. It is why these capabilities ship with a 600s timeout sized for a 30-second take instead of one extrapolated from the cheapest probe.