Seedance 2.5 timestamp-level editing: no timestamp field on Sume

Sume does not document a timestamp field for Seedance 2.5, so you cannot send "change 00:04 to 00:07" as a parameter. ByteDance describes timestamp-level control for targeted audio and video editing in its launch post; the Sume video request lists model, prompt, duration, resolution, aspect_ratio, size, frame_images, input_references, generate_audio, seed, callback_url and provider, and none of them is a time range.

That gap is worth stating plainly because it changes how you plan a workflow. If your process depends on touching only a few seconds of a finished clip, you should assume that on Sume you regenerate or recompose, not patch.

What did the vendor say about timestamp editing?

The "Introducing Seedance 2.5" page says the model offers timestamp-level control for targeted editing of audio and video, and lists green screen, camera perspective and reference-based editing as improved. It presents these as model capabilities in the Jimeng and Doubao Pro apps.

The page does not give an API field name for them in the text I fetched, so I cannot tell you how another provider exposes it.

Which request fields does Sume accept?

The table below is taken from the Sume request parameter list. Anything outside it is rejected by the strict schema rather than ignored, as covered in the strict-schema post.

Reading the table, the closest thing to temporal control is frame_images. A first_frame pins how a clip opens and a last_frame pins where it lands, which lets you design a clip as a bridge between two known images. That is useful for transitions, but it is control over the ends, not over an arbitrary moment in the middle.

How do I get a similar result with what exists?

Three workable patterns. First, describe timing in the prompt ("at the halfway point the camera cuts to the logo"); this is a request, not a guarantee. Second, generate two shorter clips and join them: set the last frame of clip A as the first frame of clip B using frame_images. Third, regenerate the whole clip with a changed prompt and the same references.

For a clip longer than a single generation, see extending past 30 seconds.

A concrete example helps. Say you have a ten second product clip where the logo should appear near the end. Rather than asking for an edit at a timestamp, generate the first six seconds with the product on a desk, then generate a second clip that begins from a still of the final frame and ends on the logo card, and join the two in your editor or in Sume Timeline. You get the beat you wanted, and each piece stays within the supported limits.

The trade-off is that two generations cost more than one and the seam can show. Match lighting and camera direction in both prompts, and use the same references so the subject looks the same. If the seam is visible, trim a few frames at the join or add a short cross-dissolve.

Another option is to accept regeneration. Keep the references and the prompt, change one sentence about the timing, and generate again at 480p to compare before paying for 1080p.

Will this change?

Possibly, but this page reflects the docs as of 2026-10-02. Check GET /v1/videos/models and the video generation docs rather than assuming. If a field is added it will show in the model entry, for example under supported_input_references.

Whatever you choose, keep a record of the prompt, references and duration for each version, so you can reproduce a clip you liked.