Claymation stop-motion look: a Gemini Omni prompt that holds

To get a claymation look from Gemini Omni, describe the material before the action: 'hand-sculpted clay figures with visible thumbprints, felt and cardboard set, lit like a miniature studio', then add a cue for the stepped motion. Omni's own guide shows that style and camera cues can be written as plain text (read 2026-10-07), and on Sume a 6-second 720p take is $0.75.

No model will give you real 12-frames-per-second stepping on demand; you are asking for a look, and you should judge it by eye.

Words that carry the look

Material and scale do more than the word 'claymation' alone. Combine them.

• Material: 'plasticine, visible fingerprints, slightly uneven surfaces'.
• Scale: 'tiny set, miniature props, tilt-shift depth of field'.
• Motion: 'slightly jerky, stepped movement, a character moves a little at a time'.
• Light: 'warm small studio lamps, soft shadows on a table'.
• Sound: 'soft clicks and clay squelches, light xylophone. No dialogue.'

A request you can run

The prompt below keeps to one character and one prop, which helps coherence in a short clip.

curl -X POST https://api.sume.com/v1/videos \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: clay-fox-1" \
  -d '{
    "model": "gemini-omni-flash-1.1",
    "prompt": "Claymation stop-motion look. A small plasticine fox with visible thumbprints carries a paper umbrella across a felt meadow on a tiny table set, slightly jerky stepped movement, tilt-shift depth of field, warm miniature studio lamps, single continuous shot. Audio: soft clay squelches, light xylophone. No dialogue.",
    "duration": 6,
    "resolution": "720p",
    "aspect_ratio": "16:9"
  }'

What a trial run costs

Style prompts need several takes, so cost per take sets how many you can afford. Sume bills list times 1.25 per second, rounded up to the cent per job.

Judging the result

Look at surfaces first. If the clay looks glossy like 3D render plastic, add 'matte, fingerprints' and remove 'cinematic'. If the motion is smooth like normal animation, the stepped cue was ignored; try 'choppy' and lower the action count to one movement. Keep the sound line short, since the model will fill any gap it finds. When a take is nearly right, change one phrase per attempt, as in the related post about 360p iteration, so you learn which word moved the result.

When to give up on the look

Some prompts never reach the style. If three takes in a row look like polished 3D, switch the method: generate a still image in the style first with an image model, then use it as the first frame so the video begins in the right look. The model tends to continue the look of the frame it starts from. That adds the cost of the image, which the Sume catalog lists, but it usually saves several video takes. The still is also easy to check before you spend on the clip.