One prompt across video models: what to rewrite on Sume
A prompt that works on one Sume video row may need edits on another: Omni uses IMAGE_REF tokens, Kling takes no references, Veo is text-only. Read 2026-10-13.
Reuse the scene text, but rewrite the parts tied to inputs. Three differences matter on Sume: Omni addresses references by token in the prompt, kling-3 rejects reference inputs, and the Veo rows accept a prompt and nothing else. A prompt sent to a row that cannot honor its inputs fails with 400 unsupported_capability; it is not silently reduced.
What to change per row
Keep the scene, subject and camera wording stable and vary only these items.
| Row | Change |
|---|---|
gemini-omni-flash-1.1 | Refer to inputs as <IMAGE_REF_0>, <VIDEO_REF_0> (counted from zero) |
kling-3 | Drop references; supply a first frame instead |
veo-3.1 rows | Text only; drop every image or clip |
wan-3.0, H3 rows | References go in input_references; no tokens documented |
| Any row | Clamp duration to the row range |
A safe test loop
Start with a 4 or 5 second clip at the lowest tier per row. Compare only the things you changed on purpose. Once a prompt passes on two rows, widen it to the long version. This also keeps the cost of a mismatch low, since each failed request fails before generation.
Sources
Related posts
More in Developers
- Playwright 1.64 lock and --shuffle for a Sume API test suite
Playwright 1.64 adds a describe lock and --shuffle. Use them to test Sume video calls: free 400 checks run in parallel, one paid test holds a credits lock.
- Python preflight: price a video batch before you submit to Sume
A short Python function prices Seedance 2.5 and Wan 3.0 jobs at list x 1.25 and compares the total with your balance before any submit. Example: $99.36.
- Python script: filter Sume video models by length, size and ratio
A short Python script reads GET /v1/videos/models and keeps the rows that fit your seconds, resolution and aspect ratio. Run it before you pick an id.
- Qwen Image 2.1 rate limits: 120 RPM Turbo, 20 RPM Pro, and batching
Alibaba caps qwen-image-2.1-turbo at 120 requests per minute and Pro at 20. Here is how to size a batch and where Sume's async mode fits.
Written by Sume