AI video generator for long videos: 30 seconds per pass, then assemble

For a long AI video, plan in short passes and assemble them. ByteDance's own page says Seedance 2.5 generates up to 30 seconds in one pass and supports multi-round extension toward multi-minute videos. On Sume, the Seedance 2.5 entry in the video catalog is 4 to 30 seconds per job, so a long piece is a set of jobs joined on Timeline 1.0.

What each side states

Keep the vendor's claim apart from what Sume exposes. Sume's docs describe what you can call today.

A plan for a five-minute piece

Write the script in scenes of 10 to 30 seconds. Generate each scene as its own job, using the last frame of one scene as the first-frame image of the next so the look carries over; POST /v1/videos accepts frame_images. Add one audio track as the spine, then place the clips on Timeline 1.0 against it.

Timeline 1.0 also has a /plan call, so look at the cut list before you pay to render. Because jobs run asynchronously, poll each one rather than holding a connection open.

What not to promise

Sume does not document a multi-minute single-call extension for Seedance 2.5, so do not plan around one. Expect drift in faces and props across scenes, review each join, and regenerate only the scene that broke.

Cost and time

Sume's docs do not give one price for a long piece, because it depends on the model, resolution and number of scenes. Read pricing_skus from GET /v1/videos/models for each model, multiply by your total seconds, and render a draft first if you can. Test a single scene at lower resolution to learn the time per scene, then budget the full set from that number.

Where Timeline fits

Timeline 1.0 is the piece that makes length possible on Sume. The audio track can run up to 1800 seconds and the video list takes 1 to 200 slots, so the limit on a finished piece is closer to half an hour than to one clip. That is a ceiling from the docs, not a promise that every long edit will look good.