Vision Settings Guide

What the vision models do here

Grout's vision models turn handwritten notes into editable text directly on the canvas, and the same models will answer questions about what an image shows. QWEN and Llama family models are among those available.

Local or cloud

• Local vision models run entirely on your machine, so nothing about the image leaves it. This is the path to use for student work and for answer sheets.
• The optional Nexus AI module reaches 300+ models on Grout's cloud GPUs when you want one your machine cannot hold. It needs an account, a connection and Merits, and the image leaves the machine.
• Larger vision models read untidy handwriting better; smaller ones answer faster. Try the small one first and only move up when it actually fails.
• Check the model's size against your free memory before downloading — the system requirements guide has the figures.

Max tokens

Max tokens caps how long the model's answer can be. It does not make an answer better, only longer — and a longer answer takes proportionally longer to generate.

• Default: 500.
• Range: 100 to 1000.
• 100–200 when you want a short label, a single figure or a one-line answer.
• 500 for a normal explanation, which is why it is the default.
• 800–1000 only when you genuinely want a long, fully worked explanation.

Temperature

Temperature controls how much the model varies its wording. Low is repeatable and literal; high is inventive, which is the last thing you want when transcribing someone's writing.

• Default: 0.7.
• Range: 0.0 to 1.0.
• 0.0–0.3 for reading text off a page, pulling out figures, or marking against a key.
• 0.3–0.5 for factual explanation where the wording can breathe a little.
• 0.7–0.9 for brainstorming, alternative phrasings and creative writing.

If the output drifts away from the page

A transcription that reads fluently but does not match the page is almost always temperature set too high. Drop it towards zero before you blame the model or the photograph.

Taking a picture the model can read

• Light the page evenly and keep your own shadow off it.
• Shoot square to the page rather than at an angle.
• Crop tightly to the writing so the model is not also reading your desk.
• Straighten a rotated page before you analyse it.
• Raise the contrast on faint pencil instead of raising the temperature.

The tools that help

Crop, rotate, fit to canvas, and the brightness and contrast adjustments all live on the canvas, so you can clean an image up in place and run the analysis again without leaving the project.

What you can ask for

• Explain image — a description of what the picture shows and what it appears to mean.
• Extract data — text and figures lifted out of the image as something you can edit.
• Custom prompt — your own question about the image, in your own words.

Writing a good image prompt

Say what the image is and what you want back. "This is a photo of a student's handwritten physics answer; transcribe it exactly, keeping the line breaks" beats "read this" every time. Save the ones that work as custom prompts so you are not rewriting them next week.

Starting points

• Transcription: the smallest vision model that copes, temperature near 0.0, max tokens sized to the page.
• Explaining a diagram: a mid-sized model, temperature around 0.4, max tokens 500.
• Marking an answer sheet: temperature 0.0–0.2, max tokens 500, with your mark scheme in the prompt.
• Ideas from a sketch: temperature 0.7 or above, max tokens 500.

Changing one thing at a time

When a result is wrong, change the photograph, the prompt, the temperature or the model — but only one of them per attempt. Changing two means you learn nothing from the result.