Wan 3.0 instruction-based editing vs Sume's Omni edit: what each takes

Instruction-based video editing means you hand over a clip and say in words what to change. Alibaba lists it for Wan 3.0 and ByteDance lists timestamp-level editing for Seedance 2.5, but on Sume the one documented edit mode is video_url on gemini-omni-flash-1.1, at $0.125 per second at 720p.

Edit features, side by side

Each row is cited to the page named in the source column. Vendor features are what the vendor says its own model does; the Sume column is what the Sume docs say you can send.

Where each claim comes from

The vendor sources are the Wan 3.0 repository, ByteDance Seed's Seedance 2.5 post and Google's Gemini Omni 1.1 Flash post, all read 2026-10-05. For Seedance, the post says the API is coming soon via BytePlus ModelArk, so a vendor feature is not the same as a callable one.

What the Sume edit takes

On Sume, an edit request has one required media field, video_url, and a prompt. The Sume docs say video_url is the edit source, not a reference. It cannot be sent together with image_url, end_image_url or the reference_*_urls fields. Native synced audio is always on for this model.

The output keeps the source framing: aspect_ratio is rejected on edit, and the source clip sets the length. The docs give the 360p to 4K resolution range, 720p by default.

Price check

Because Wan 3.0 and Omni Flash 1.1 share the Sume 720p rate of $0.125 a second, the choice is about capability, not cost. A 10 second Omni edit is $1.25; a 10 second Wan 3.0 regeneration is $1.25. At 1080p they split: Omni at $0.1875 a second against Wan at $0.25.

A short decision rule

Three questions settle it. Do you need to keep the original clip? Then edit with Omni. Do you need a clip longer than 10 seconds from a prompt? Then Wan 3.0, up to 30 seconds. Do you need box or timestamp control? Neither is documented on Sume, so plan to do that step in your own editor, or check the vendor's own surface.

Prompt shape

Write edit prompts as a change plus a constraint, as in the Sume docs example: "Replace the bottle with an apple. Keep everything else the same." Vague prompts like "make it better" give the model nothing to hold fixed.

Test the instruction, not the promise

Instruction-based editing is easy to over-trust because the instruction reads like a promise. The vendor pages list the capability; they do not publish pass rates in the sections we read. The sensible test is your own: take three clips you know well, make the same instruction on each, and write down what moved that you did not ask to move.

Where the edit changes lighting, motion or faces, check continuity at the cut points. Native audio is always on for Omni Flash 1.1 on Sume, so also listen: an edit that changes the scene can change the sound bed.

Seedance 2.5 on Sume

Seedance 2.5 deserves a note. ByteDance's post lists up to 30 images, 10 video clips and 10 audio clips per pass, timestamp-level editing, green-screen background replacement and camera perspective change (read 2026-10-05). Sume's Seedance 2.5 row accepts 4 to 30 seconds at 480p, 720p and 1080p. Sume sets its own reference caps per model, which you read from capabilities in the catalog. Features on the vendor's own surface and fields on Sume are two different lists, and only the second is something you can call here.

Sources: Video Router docs for the Sume fields and limits.