Integration guides · 2026-09-22
The Focus tool on Perceptron Mk1: fine-grained analysis by zooming into a region
Covers how the Focus tool, described in Perceptron's official documentation, lets Mk1 zoom into a specific region of an image through a tool call, and when it's useful.
Focus turns a general description into a detailed inspection
According to Perceptron's official documentation, the Focus feature lets Mk1 zoom into specific regions of an image through tool calls to produce fine-grained answers. That means the model can inspect a small region relevant to your question at a separate resolution, instead of evaluating the whole image with one general overview.
This capability produces a more accurate result than a general image description in scenarios where a small detail in a wide image matters — small text on a sign, a field in the corner of a form, an object in the background of a scene.
When it kicks in, how it's triggered
Focus is a Perceptron-specific field, not OpenAI's standard `tools`/`tool_choice` mechanism: you can request it explicitly by sending `internal_tools.focus: true` inside the request's `vision_config`, or leave it to the model's own judgment. Phrasing your question to point at a specific region rather than a general description (asking 'what does the sign in the top-right corner say,' for example) increases the likelihood of the model invoking Focus.
The model also accepts video input and supports tasks like video Q&A, finding event moments with timestamps (video clipping), and video-based in-context learning; Focus's image-region zoom logic is limited to static images, and video uses a separate mechanism.
- Focus is an image-region zoom capability that operates through a tool call.
- Phrasing your question to point at a specific region increases the chance Focus gets triggered.
- Video input brings separate capabilities into play (Q&A, event finding, in-context learning).
Verify format and size limits
Focus's full parameter set and the supported image format/size limits aren't covered in detail in Perceptron's official documentation available at the time of writing; verify these details against the current API reference or the model page before going to production.
Frequently asked questions
Do I need to explicitly invoke Focus in my request?
Focus operates as a tool call; the model can also trigger it on its own judgment if it decides your question needs detail. Phrasing your question to point at a specific region increases that likelihood.
Can Perceptron Mk1 zoom into a region on video input too?
Video has separate capabilities listed, such as video Q&A, event-moment finding, and in-context learning; check the model page for whether that is identical to Focus's static-image zoom logic.