How many reference images can I attach to one prompt?

Because its references are typed. Pro distinguishes an object reference — a product, a garment, a prop it should reproduce faithfully — from a character reference, a face it should keep consistent across scenes. The budget is 6 objects plus 5 characters, 11 in total, and the two pools do not trade against each other.

The other image models treat every reference as generic visual context, which is why they accept more of them but hold identity less tightly. If you are building a character who has to look the same across a series, Pro's smaller, typed budget is the better tool.

Veo treats them as two different tasks. A first frame means *animate this exact picture*, and the model is bound to those pixels for frame one. Reference images mean *this is what the subject looks like, now compose a shot* — the model chooses the framing itself. The API exposes them as alternatives, and asking for both is rejected before generation.

Gemini Omni Flash takes the opposite approach: roles are declared as tags inside the prompt text rather than as separate API fields, so a start frame and up to nine additional references can travel in the same request, ten images in all.

In the studio, an upload or any image already in your history — the second option matters, because reusing a generated character as a reference is how consistency across a series is built without re-uploading anything.

Over MCP the same slots accept three forms: a job id from an earlier generation, a public http(s) URL, or an inline data:image/...;base64, payload up to 10 MB. A bare local file path will not work — the server cannot read your disk — which is the most common mistake when wiring an agent up for the first time.

References never change the price. What you pay is set by the model, resolution and count of outputs; attaching eleven photos costs exactly the same as attaching none.