[Feature]: Google Imagen models support

[jaimeiniesta]

Feature Type

Provider feature extension

Provider (if applicable)

google

Description

When I run ReqLLM.Images.supported_models() I see the imagen models listed. However these fail with a message like models/imagen-4.0-generate-001 is not found for API version v1beta, or is not supported for generateContent. Call ListModels to see the list of available models and their supported methods.

Gemini models like gemini-2.5-flash-image and gemini-3-pro-image-preview work fine, but imagen models do not work.

Example Usage

iex(2)> ReqLLM.Images.supported_models
["google:gemini-2.5-flash-image", "google:imagen-4.0-generate-preview-06-06",
 "google:imagen-4.0-ultra-generate-001",
 "google:imagen-4.0-ultra-generate-preview-06-06",
 "google:gemini-3-pro-image-preview", "google:imagen-4.0-generate-001",
 "google:gemini-2.0-flash-exp-image-generation",
 "google:imagen-4.0-fast-generate-001", "openai:dall-e-3", "openai:gpt-image-1",
 "openai:gpt-image-1.5", "openai:gpt-image-1-mini",
 "openai:chatgpt-image-latest", "openai:dall-e-2",
 "openrouter:google/gemini-2.5-flash-image-preview",
 "openrouter:openai/gpt-5-image-mini",
 "openrouter:google/gemini-2.5-flash-image", "openrouter:openai/gpt-5-image",
 "xai:grok-2-image-1212"]

Can you contribute this?

Yes, I can submit a PR

[vkryukov]

Investigation Summary

I've identified the root cause of why Imagen models don't work.

The Problem

The Google provider at lib/req_llm/providers/google.ex:444 hardcodes the API endpoint for all image models:

url: "/models/#{model.id}:generateContent"

However, Imagen models and Gemini image models use different API methods:

Model Type	API Method	Request Body Format
Gemini (gemini-2.5-flash-image, etc.)	:generateContent	{"contents": [{"parts": [{"text": "..."}], "role": "user"}]}
Imagen (imagen-4.0-generate-001, etc.)	:predict	{"instances": [{"prompt": "..."}], "parameters": {"sampleCount": N}}

Evidence

1. The error message confirms this: "models/imagen-4.0-generate-001 is not found for API version v1beta, or is not supported for generateContent"

2. Google's Imagen documentation shows the correct endpoint is :predict with a different request body structure.

What Needs to Change

To fix this, the Google provider needs to:

1. Detect Imagen models - Check if the model ID contains "imagen" (or check supported_generation_methods in model metadata)

2. Use the correct endpoint - Route to :predict for Imagen models instead of :generateContent

3. Use the correct request body format - Imagen requires:

   {
     "instances": [{"prompt": "your text prompt"}],
     "parameters": {"sampleCount": 1}
   }

4. Parse the different response format - Imagen returns responses in a different structure than Gemini's generateContent

Key Files to Modify

• lib/req_llm/providers/google.ex:397-463 - prepare_request(:image, ...) needs model-type branching
• lib/req_llm/providers/google.ex:797-825 - encode_image_body/1 needs a separate encoder for Imagen
• Response decoder also needs to handle Imagen's response format

References

• Generate images using Imagen | Gemini API
• Image generation API | Vertex AI