Answers

Do seeds actually reproduce the same AI image?

By BananaBanana TeamPublished Last checked

Short answer

Partly. On the image models a seed is a real control: reusing one with the same prompt and settings keeps composition close, which is how you iterate on wording without losing the frame. On Gemini Omni Flash video it is accepted but ineffective — two runs with an identical seed and prompt came out visibly different in our tests. For consistency across generations, reference images beat seeds.

Image models
Seed accepted, keeps composition close across runs
Gemini Omni Flash
Seed accepted, no measurable reproducibility
Measured
Two identical-seed runs scored ~0.60 structural similarity
Stronger tool
Reference images and multi-turn editing
Cost of a seed
$0.00 — it is a free parameter

Facts on this page were checked against the live platform on .

01In detail

What a seed does and does not promise

A seed fixes the random starting point of a generation. Hold everything else constant and the result stays in the same neighbourhood — same rough composition, same colour direction — which makes it genuinely useful for A/B-ing prompt wording: change three words, keep the seed, and the difference you see is the words.

What it does not promise is bit-identical output. These are hosted models behind a serving stack that changes underneath you; a seed is a strong hint, not a hash. Anyone promising exact reproducibility across days on a hosted frontier model is overselling.

Why is the video seed ineffective?

We tested it rather than assuming. The Interactions API accepts a seed field on Gemini Omni Flash without complaint, but two runs with the same seed and the same prompt produced clips that differ in framing and staging — a structural similarity around 0.60, which is nowhere near a reproduction.

The honest reading is that the field is plumbed but not honoured end to end in this model. So do not build a workflow that assumes a seed will bring a shot back; store the generated clip instead, and edit from it.

What to use when you need consistency

Reference images. Feeding the model a face or a product and asking for a new scene holds identity far better than any seed. On Nano Banana Pro these are typed — up to 6 object and 5 character references — which is the strongest consistency tool available here.

Multi-turn editing. Instead of re-rolling a prompt, edit the image you already have: the model receives the conversation history, so *same shot, warmer light* changes the light and leaves the shot alone. No re-upload is needed.

Conversational video editing. On Omni Flash the same idea applies to clips: ask for a change to an existing generation and the character and lighting survive the edit in a way a re-roll with a fixed seed will not.

02Go deeper

Where this is documented

03Guides

Longer reads on the same thing

04Also asked

More on this question

Where do I set a seed?

In the studio parameters panel, next to the dice button that generates a random one. Over MCP it is an optional argument on the image tools.

Does the same seed work across models?

No. A seed is meaningful only within one model; the same number on Nano Banana 2 and Nano Banana Pro is unrelated.

Will a fixed seed get me past a content filter?

No. The output classifier judges the rendered pixels each time, so a filtered result may still pass on a retry — but the seed is not what decides it.

Is there any charge for using a seed?

None. Seeds, aspect ratios, negative prompts and reference images are all free; price depends on model, resolution, duration and output count.

05Related

Questions next door

Try it on your own prompt

New accounts start with $0.20 of balance — no card, nothing expires.