Seed Value
Whether you can fix randomness with seeds in Nano Banana Pro 2—and what the Gemini API docs imply for reproducibility.
Seeds in the web UI
Nano Banana Pro 2 does not expose a seed field. You cannot type a seed or copy one from metadata the way some desktop tools allow.
Practical implication: treat each generation as stochastic. For “almost the same” results, keep model, aspect ratio, resolution, reference images, and prompt fixed and run again—or use multiple images per request and pick the best frame.
What Google’s stack does elsewhere
The Gemini image generation documentation focuses on prompting, imageConfig (aspect ratio, image size), tools (e.g. Google Search), and multi-turn history—not end-user seed picking. For API integrations, reproducibility still depends on provider behavior; always verify in your own pipeline.
Workflow substitutes
| Goal | Approach |
|---|---|
| A/B one word | Lock everything else; change a single phrase; same number of images per run. |
| Series coherence | Attach the previous output as a reference for the next edit (character 360°, style lock). |
| Batch review | Use 2–4 images per prompt to sample variance without manual seed sweeps. |
Related
- Chaos and variety for exploration without chaos/seed controls.
- Cross-image character consistency guide for identity-locked series.