I’ve tested both gemini-3-pro-image-preview and gemini-3-pro-preview with candidateCount: 2 and always get this error:
"Multiple candidates is not enabled for this model" — INVALID_ARGUMENT (400)
This appears to be a current platform limitation — the API simply does not support returning multiple image candidates from a single prompt/seed.
The usual suggestion of “just fire parallel requests” does not solve the problem, because parallel requests will produce different images each time — they don’t share the same generation context/seed. The whole point of candidateCount is to get multiple variations from the same generation pass, so you can let the user pick the best one while keeping consistency.
My ask:
-
Is there a workaround to get true multi-candidate image generation (same seed, multiple outputs)?
-
Is this feature planned or on the roadmap?
-
Is there any way to at least fix the seed so parallel requests produce consistent/comparable variations?
Using @google/genai ^1.40.0.