Make a music video from your own footage with AI: Sume restyle
sume-restyle redoes a clip you own with the same motion and cuts, in chunks of up to 30 seconds. What to supply, what stays, and the audio rule.
If you already filmed a performance and want it to look like something else, the catalog Format for that is sume-restyle. It restyles a clip you own, keeping the same motion and the same cuts. It is not a text-to-video tool: you bring the footage, and the Format changes how it looks.
How the Format works
According to its recipe in the Sume repository, sume-restyle sends each chunk of up to 30 seconds through one Seedance 2.5 reference-to-video call, with generate_audio off. It then assembles the chunks on a timeline and runs a drift check on the result. The 30 second chunk matches the Seedance 2.5 per-pass length in the model catalog.
| Stage | What happens | What you control |
|---|---|---|
| Input | Your clip, which you own | The footage and the target look |
| Generate | One Seedance 2.5 call per chunk of up to 30 s | The instruction |
| Audio | generate_audio off | Use your own track, only if you own it |
| Assemble | Timeline joins the chunks | Review each seam |
| Check | Drift check against the source | Rerun a chunk that drifts |
The audio rule
Music is the part of a music video most likely to cause trouble. The Format does not generate audio for the restyle, and the recipe says to keep audio only if the user owns it. Use a track you wrote, or a track you are licensed to use, and put it on the final timeline yourself.
If you need a new song, the catalog has a music router. The model ids there are sume/music-auto, lyria-3.5 and lyria-3-pro, and there is no duration field, so steer length in the prompt.
How to brief it
Describe the target look in a sentence and name what must not change, for example the singer's position and the cut timing. Because the Format keeps motion, a clean, well-lit original gives the model more to hold on to than a dark, shaky one.
Run a short test chunk first. When the look is right, run the full clip, then check the join between chunks, where drift is most likely to show.
What to expect
Restyling is not a free pass on quality. Fast motion, hands and instruments are the places where a generated look is most likely to wobble, and the drift check is a safety net, not a guarantee. Expect to rerun a chunk or two. Because each chunk is its own call, a rerun costs only that chunk and not the whole song.
Sources
Related posts
More in Formats
- No queue webhook on Sume bulk runs: count item webhooks instead
Sume bulk queues have no queue-level webhook. Put a webhook_url on each item and count terminal events to know a season is finished.
- Output schema 400 missing_items: an array node needs items
A Format output_schema with an array and no items fails with 400 and rule missing_items before anything runs. The bad schema, the fix, the violation.
- Output schema unsupported_ref: $defs must live at the schema root
A $ref to a nested $defs, an external URL or a missing name fails with unsupported_ref in a Sume output_schema. Move $defs to the root and fix the pointer.
- Output schema max_enum_values: 1,000 values per enum, then what
Sume refuses an enum of more than 1,000 values with rule max_enum_values. Use a string with a pattern or format, or split the list into two enums.
Written by Sume