Fix one detail in a finished AI video: three routes on Sume
Seedance 2.5 mentions local editing. On Sume your options are Omni edit, Recast for a person swap, or a new take. Pick by what must stay unchanged.
On Sume there are three ways to fix one detail in a finished clip: edit it with gemini-omni-flash-1.1 and video_url, swap the people with h3-max-recast, or generate a new take. None of them is a mask-based local edit of one region of the kind ByteDance describes for Seedance 2.5, so pick the route by what must stay unchanged.
ByteDance's Seedance 2.5 page describes editing and reference capabilities; Sume documents no request fields for that, and Sume's seedance-2.5 row takes video input only as a reference, not as an edit source.
Route 1: Omni edit
Send the clip as video_url with a prompt such as "Change the shirt to red. Keep everything else the same." The source sets the output length and Omni keeps native synced audio. The model is 3 to 10 seconds, so the clip has to fit that range. Expect a regenerated clip, not a patch, so check details that were not part of the edit.
Route 2: Recast for people
h3-max-recast replaces the primary people in a source video with the people in 1 to 4 reference photos, one photo for each new person. It keeps the source motion, camera, cuts and sound. The source can be 5 to 30 seconds with no single shot longer than 15 seconds, and the output is 768p (default) or 1080p.
Route 3: a new take
If the problem is the performance, not one object, regenerate with a revised prompt. Sume video has no seed, so a re-run is a new take. Keep the good takes and mark them, since you cannot recreate one by repeating the request.
Seedance 2.5 in particular accepts up to 30 seconds and reference images, videos and audio, so a corrected reference set often fixes a drift that a prompt change cannot.
| Problem | Route | Keeps | Limits |
|---|---|---|---|
| An object or color is wrong | Omni edit (video_url) | most of the clip and its audio | 3 to 10 s source, edit may shift details |
| The wrong person is in it | h3-max-recast | motion, camera, cuts, sound | 5 to 30 s, no shot over 15 s |
| The performance is wrong | new take with references | nothing, a new clip | no seed, results vary |
| Background is wrong | Omni edit (video_url) | subject and audio | prompt describes what to keep |
A quick decision rule
Ask what must stay identical. If it is the whole clip except one object, try Omni edit first, because it is the closest to a local change. If it is the choreography and only the person differs, use Recast. If nothing needs to survive, rewrite the prompt and generate again.
- Edit cost scales with the length of the clip and the resolution you choose.
- Recast is listed at $0.30 or $0.45 per second before the 1.25 multiplier, above Omni at 720p, so confirm the amount before a long source.
- Always compare the result with the source, frame by frame, around the edited area.
Sources
Related posts
More in Use cases
- Fix one typo in an AI-generated image: quote the old and new text
Send the image as the first input_references entry and name the wrong text and the right text in quotes. Ideogram 4.5 on Sume costs $0.0375 to $0.275 per try.
- Fix the last second of an Omni clip: edit it or generate a new ending
A bad last second in a 10 s Omni clip: edit the whole clip for $1.25 at 720p, or generate a new 3 s ending for $0.38. Cost and risk of each route on Sume.
- Six flavor variants of one can: recolor and rename with Ideogram 4.5
One can render becomes six SKUs by editing color and flavor name in one Ideogram 4.5 call each: $0.225 at low, $0.45 at medium on Sume. Prompt and checks.
- Floor-plan sketch to a furnished room render on GPT Image 2.5
Turn a floor-plan sketch into a furnished room render with GPT Image 2.5 at 16:9. The model does not measure, so name sizes in the prompt. Code and prices.
Written by Sume