Skip to content
TC
English
Side Projects

Netron OCR

Turn Netron screenshots into model graph JSON.

Upload a picture or screenshot of an ONNX model open in Netron, recover the nodes, tensor shapes, and connections, then correct the result as structured graph JSON.

Input

Netron screenshot

Output

Graph JSON

Use

LLM model rebuild

The recovered graph can be inspected and corrected directly over the source screenshot.

Why this exists

Why screenshots alone are not enough for model reconstruction.

Why Screenshots Are Not Enough

I built Netron OCR after finding that frontier LLMs, including GPT-5.5-class systems, are still bad at recreating an ONNX model directly from a screenshot. The same models do much better when the graph is described as text or JSON.

Manually describing every operator, tensor shape, and edge in a Netron graph is slow. Netron OCR closes that gap by converting the screenshot into structured annotations that can be corrected in place.

Prediction and correction happen directly on top of the original Netron image.

From Netron Image To LLM Handoff

Upload a picture or screenshot of an ONNX graph in Netron, let the pipeline recover the visual structure, then fix any node labels, tensor shapes, or connections in the editor.

The resulting JSON describes the graph in a form an LLM can read. Instead of asking the model to understand pixels, you can give it the recovered graph description and ask it to reproduce the ONNX model from that structure.

The editor keeps the screenshot, graph overlay, inspector, and JSON view in one workflow.

What it does

From a Netron image to a graph description an LLM can use.

Image to graph JSON

Recover nodes, tensor shapes, graph inputs, graph outputs, and connections from Netron-style model screenshots.

Editable recovery

Correct labels, boxes, edges, and shape attachments directly over the source image before using the structured result.

ONNX-focused workflow

The annotations are shaped around operator-level graph reconstruction rather than generic document OCR.

LLM handoff

Feed the corrected JSON description to an LLM so it can rebuild the model from structure instead of guessing from pixels.