How it works — Image Analysis
Upload any image and ask NuBench AI to analyze, describe, or extract data from it — powered by GPT-4o and Claude 3 vision capabilities. Images are classified, DLP-scanned, and governed by the same pipeline as every other NuBench request.
Use the paperclip icon in the chat input to attach one or more images to your message. Supported formats are PNG, JPEG, GIF, and WebP up to 5 MB per image.
Images are uploaded over an encrypted connection and associated with your conversation record. They are stored at the same classification tier as the conversation and subject to the same DLP scanning as text content.
When an image is attached, NuBench automatically routes the request to a vision-capable model. The model picker shows which models support vision — GPT-4o and Claude 3 Sonnet, Opus, and Haiku are eligible by default.
Model eligibility for vision is configured in the model registry and subject to the same classification constraints as text-only models. PHI-classified conversations are restricted to BAA-eligible vision models.
Vision models can describe the contents of an image in natural language, extract text and structured data visible in the image, answer questions about specific regions or elements, and compare multiple images in a single turn.
Performance varies by model and image content. For document-heavy images like scanned pages or screenshots, Claude 3 Opus typically produces more accurate text extraction. For photographic images, GPT-4o typically produces richer descriptive output.
Images are passed through the DLP pipeline before being sent to the model. The DLP engine applies optical character recognition to extract visible text and scans it against your active DLP rules.
If the image contains content that would trigger a DLP block in text form — a PII pattern, a restricted keyword, or a content type you have blocked — the request is stopped before the model sees the image. The audit trail records the block event.
Upload a scanned contract, invoice, or form and ask NuBench to extract key fields, summarize the terms, or flag unusual clauses. The model reads the document as a human would, without requiring a PDF parser or OCR pipeline.
Paste a screenshot of a chart, dashboard, or financial table and ask NuBench to describe the trends, extract the underlying numbers, or draft a written summary for a report.
Share a screenshot of an interface — a web page, a mobile app, or a design mockup — and ask NuBench for feedback on layout, accessibility, consistency, or user flow.
Upload an architecture diagram, process flow, network map, or org chart and ask NuBench to explain what it depicts, identify gaps, or suggest improvements.