Image to Text¶

Image to Text is a first-class Bro mode and the priority use of Gemma 3's multimodal capability.

It accepts image input and returns text. It does not generate images.

Requirements¶

Image analysis requires both:

the configured Gemma 3 4B IT GGUF; and
a matching mmproj multimodal projector GGUF.

See Local Model for projector discovery and configuration.

Supported image types¶

PNG
JPG
JPEG
WEBP

Multiple images can be uploaded when the selected task requires comparison.

👁️ Vision Controls¶

A single six-control row provides:

Control	Purpose
Vision Task	Selects the image-understanding operation.
Image Detail	Controls response depth.
Response Format	Uses Bro's standard bounded output formats.
Response Language	Uses the shared bounded human-language list.
Preserve Layout	Requests preservation of visible document/layout structure.
Include Visible Text	Requests explicit transcription of visible text.

Vision tasks¶

Extract Visible Text
Describe Image
Answer Questions
Analyze Screenshot
Analyze Chart
Analyze Diagram
Extract Structured Data
Compare Images

Extract Visible Text is the primary OCR/Image-to-Text workflow.

🎚️ Inference Settings¶

Vision uses the same generation controls as text generation:

Row 1¶

Temperature
Top-P
Top-K
Repeat Penalty
Repeat Window

Row 2¶

Presence Penalty
Frequency Penalty
Random Seed
Max Tokens

One bottom Reset resets both rows.

⚙️ Runtime Settings¶

Context Window
CPU Threads
Batch Size
Micro Batch Size
Projector Device

Projector Device is a bounded CPU / GPU selectbox.

System Instructions¶

Image to Text retains the standard 🖥️ System Instructions surface so a saved vision-oriented prompt can be applied without creating a separate prompt-storage mechanism.

Request flow¶

upload image(s)
    |
    +-- select Vision Task
    |
    +-- optional Image Request
    |
    +-- build vision message
    |
    +-- MTMDChatHandler + mmproj
    |
    +-- Gemma 3
    |
    +-- streamed text response

Capability gating¶

If the multimodal projector is missing or incompatible, Bro warns the user instead of attempting to send images to a text-only runtime.

Relationship to Document Q&A¶

Document Q&A can reuse this vision pipeline when OCR is enabled and a PDF page lacks usable native text. This makes Image-to-Text a shared runtime capability rather than an isolated demonstration.