Is there a seed to keep Shorts characters the same across episodes?
No. No video model in Sume's catalog accepts a seed, and the docs say a seed is not deterministic on all providers. Use references and first frames instead.

No. In the Sume repo's catalog no video model accepts a seed, and the Videos docs note that the seed field is not deterministic on all providers. Consistency across a series comes from what you send in, not from a number.
That changes how to plan a season: treat every clip as a fresh draw and spend your effort on references.
What to use instead
There are three levers in the request. input_references gives image or video guidance. frame_images pins a first or last frame for image-to-video. The prompt itself repeats a fixed description of the cast and set. Pick the first for look, the second for continuity of a cut, and always do the third.
| Tool | Field | Holds | Does not hold |
|---|---|---|---|
| Reference images | input_references | Look, wardrobe | Exact pose |
| First frame | frame_images (first_frame) | The opening shot | The rest of the clip |
| Last frame | frame_images (last_frame) | The closing shot | Path between frames |
| Fixed prompt block | prompt | Names, setting | Face identity alone |
A season routine
Write a one-paragraph character card and paste it unchanged into every prompt. Keep two or three approved stills per character. Re-roll beats, not episodes, so a failure costs seconds: at Wan 3.0 720p a 30-second beat is $3.75.
Read the Videos page for the seed flag on the model you pick; it is listed per model in the catalog response.
- Approve stills first, then animate.
- Keep the character card in version control.
- Never promise a client the same face twice; promise the same references.
Writing a character card that works
A card is a fixed paragraph that you paste into every prompt for a character. Keep it concrete: age range, hair, clothing in two colors, one distinct accessory, and the camera distance. Avoid adjectives such as 'cinematic' in the card; those belong in the scene part of the prompt.
Pair the card with approved stills. Generate the stills first with the images API, choose the three best, and use them as input_references for every video request in the season. This way the video model sees the same cast each time, and your prompt only has to describe what happens.
Expect drift anyway. Over a long season the cast will change a little. Decide in advance what you will accept, and re-roll beats that cross the line instead of chasing a perfect match. The cost of re-rolling a 10-second beat on Omni Flash at 720p is $1.25.
One more habit: log the full request body for every kept clip. Without a seed you cannot rebuild a take from a number, but you can rebuild the conditions, and a body with the prompt, references and model id is the closest thing to a recipe.
Sources
Related posts
More in Models
- Silent clips after Sora: which Sume models take generate_audio false
Gemini Omni Flash 1.1 rejects generate_audio false; Kling 3 prices audio on at $0.21 a second against $0.14 off; recast and motion transfer keep source sound.
- A 60-second Seedance 2.5 piece: two 30 s jobs and one last frame
Together lists multi-round extension for Seedance 2.5; Sume documents none. Build 60 s as two 30 s jobs and a last-frame handoff: $34.68 at 720p, $7.50 on Wan.
- Square 1:1 ad video: which Sume models accept it, and 4:5
Seedance, Kling, Wan and MiniMax accept 1:1 on Sume; Gemini Omni takes only 16:9 and 9:16. No video model lists 4:5, so Feed needs a crop. Table and steps.
- Should a Shorts season use sume/auto or a pinned video model?
Pin one model per season so episodes look alike. Use sume/auto for pilots and one-offs where a consistent look is not required.
Written by Sume