Pixel art and isometric video with Omni Flash: prompts

To get a pixel-art or isometric look from Gemini Omni Flash, name the look, the camera and what stays fixed, then draft at 360p. Example: Isometric 3D diorama of a tiny bakery, fixed isometric camera, soft shadows, slow loop. On Sume a 6-second 720p draft is $0.75 and a 360p draft is $0.225. Google's Omni page, read 2026-10-08, lists no style controls beyond the prompt, and no temperature, top_p or system instructions.

That means style is a matter of wording, references and iteration. The model cannot be told to be more or less random.

Wording that tends to hold a look

These are prompting habits, not guarantees. Test each at 360p.

• Pixel art: 16-bit pixel art, limited palette, hard pixel edges, no anti-aliasing, side-scrolling camera. Mention a frame rate feel such as animated at 8 frames per second if you want the stepped look.
• Isometric: fixed isometric camera at 30 degrees, no perspective change, orthographic look. Add locked camera so the model does not drift.
• Give a reference image of the style. Sume's Omni row takes up to 10 reference images and the prompt calls them <IMAGE_REF_0> and so on.
• Say what must stay: Keep the same palette throughout.
• Use No scene cuts if you want one continuous shot.

What the limits mean for stylized work

Omni outputs are 16:9 or 9:16, so a square sprite sheet needs cropping later. The model makes sound on its own, so ask for chiptune, or ask for silence in words, but the track cannot be removed through the API. Native 4K is not native: Google labels 1080p and 4K as upscaled, and an upscale of hard pixel edges will soften them. For pixel art, 720p is the honest size.

Compared with Veo 3.1

Veo 3.1 clips are 4, 6 or 8 seconds, and Google's Veo page says 8 seconds are needed with reference images or at 1080p and 4K. Veo 3.1 Fast lists $0.10 per second at 720p on the pricing page, so a 6-second draft is $0.60 against $0.75 for Omni on Sume. Sume lists no Veo model. If the look depends on a style image, Omni's limit of 10 reference images is the more useful number.

A testing routine

Change one thing per draft: first the style line, then the camera line, then the palette. Twenty 360p tries cost $4.50 on Sume and show which phrase moves the look. Keep a note of the wording that worked, since there is no seed to repeat a result and no setting to make the model less random.

Submit with an Idempotency-Key header so a retry after a dropped connection returns the original job instead of a second charge; the same key with a different payload returns 409. Then poll GET /v1/jobs/:id/status and fetch GET /v1/jobs/:id/result when it finishes.