# `LlmCall.Media.Dialect.Gemini.Predict`
[🔗](https://github.com/gilbertwong96/llm_call/blob/main/lib/llm_call/media/dialect/gemini/predict.ex#L1)

Gemini Imagen image generation (`POST /v1beta/models/{model}:predict`).

Instances carry the prompt; media controls live under `parameters`. `size`,
`seed`, and `duration_s` are rejected because this endpoint does not expose
them, and `references` are rejected because the JSON body cannot carry an
upload.

`model` must be a bare model id: it is interpolated into the request path.

A filtered prediction carries `raiFilteredReason` instead of image bytes; it is
skipped while another prediction has bytes and becomes a `:provider` error when
none do. Without `includeRaiReason` in `provider_options.gemini["parameters"]`
the provider omits filtered images altogether, so an empty `predictions` array
means everything was blocked rather than that the payload was malformed.

# `validate_request`

```elixir
@spec validate_request(LlmCall.Media.Request.t(), LlmCall.Config.t()) ::
  :ok | {:error, LlmCall.Error.t()}
```

Runs the shared media validation pipeline with the codec's own checks.

---

*Consult [api-reference.md](api-reference.md) for complete listing*
