Preprocessing of hosted `gemma-4-26b-a4b-it`

Gemma
,
1

Hi, I’m looking for clarification about image preprocessing for the hosted model:

gemma-4-26b-a4b-it

I used it through the Google Generative Language API, with 256×256 RGB PNG images sent inline as base64 image/png.

I would like to know what preprocessing this hosted model applies before the vision encoder sees the image.

More specifically:

  • Does the hosted gemma-4-26b-a4b-it use the same processor as the public google/gemma-4-26B-A4B checkpoint?
  • If yes, what are the exact preprocessing details eg
    • resize behavior
    • aspect-ratio handling
    • interpolation
    • padding/cropping
    • tiling
    • vision-token/image-token budget
  • If no, is the hosted preprocessing pipeline documented anywhere?
  • Is there any way to see the processed image size, token count, or backend/model revision for reproducibility?

Feel free to ask any questions if you need more details.

Any clarification would be very helpful. Thanks!

2

Hi @legion1
The hosted gemma-4-26b-a4b-it model uses Gemma 4 vision, but its server-side image processor isn’t exposed or versioned separately in the Gemini API. The public Gemma 4 processor documentation is the reference for the model’s vision preprocessing, including image resizing, patching, and visual-token budgeting. For the detailed preprocessing behavior, please refer to the Gemma 4 Vision documentation and the Visual Guide to Gemma 4

Thanks

3

Hi @Pannaga_J

Thank you so much for clarifying my doubts.

One detail would resolve the main reproducibility issue for me: for gemma-4-26b-a4b-it requests through the Google Generative Language API where I did not explicitly input in a visual token budget, what token budget/default was used for a 256×256 PNG?

Also, when you say the public Gemma 4 processor documentation is the reference, can I interpret that as meaning the hosted preprocessing follows the same resizing/aspect-ratio/patching behavior as the public Gemma 4 processor, even though the server-side implementation itself is not separately revealed?

4

Does anyone have an honest answer to this?