An image gives you pixels. JSON gives your software the geometry.

DWG to image renders the drawing geometry onto a raster canvas and saves the result as a PNG, JPEG, or TIFF. The output is a flat grid of colored pixels. It is useful for displaying a thumbnail, embedding a preview in a PDF report, or giving a non-CAD user a visual reference. It is not useful for any downstream software processing.

DWG to JSON reads the binary DWG format and emits a structured description of every entity in the drawing: its type, coordinates in native drawing units, layer membership, color, and bounding box. The output is machine-readable data your code can query, filter, measure, and transform.

An image conversion is a one-way street. The moment the DWG is rendered to pixels, the structural information in the file is gone:

• No measurements. A pixel has no real-world size. You cannot determine that a wall is 4.5 metres long or a bolt hole is 12 mm in diameter from a PNG.
• No layer structure. Layers in a DWG file carry intent — architectural, structural, electrical. A flat image collapses all layers into one surface with no way to distinguish them.
• No entity classification. Whether a line is a wall, a road edge, or a pipe run is encoded in the layer and block names. That information does not survive in pixels.
• No text extraction. Annotations, part numbers, and dimension labels in a DWG are searchable text entities. In a rendered image they are just dark shapes on a background.
• No automation. You cannot write code that counts door entities, sums cable lengths, or filters structural elements from an image without expensive computer vision work.

The CADLens sheets[].entities opens up a category of features that are simply not possible from an image:

• Automated measurement. Calculate total polyline length (cable runs, pipe lengths, perimeter), area of closed regions, or count of circular entities (bolt holes, columns).
• Layer-based filtering. Show only the structural layer. Strip annotation layers before displaying to clients. Count entities per discipline.
• BOM and take-off generation. Identify block inserts by name, aggregate counts, and export a bill of materials — all without a human reading the drawing.
• Search and indexing. Extract all text entities from a drawing library and make them full-text searchable. Find drawings that reference a specific part number.
• Downstream export. Feed the vector data into quotation engines, GIS systems, or custom rendering pipelines that need real coordinates, not pixels.

The metadata object adds drawing-level context: units (millimetres, inches, etc.) and the overall bounding box ({ width, height }) so downstream code knows the coordinate space.

CADLens is a parser, but it also renders a PNG from the same parse run. You do not have to choose. A single POST /v1/parse call returns a job_id. When the job reaches COMPLETED, the result from GET /v1/jobs/{job_id}/result includes:

• sheets[].entities — every entity with coordinates, layer, color, and bounding box, grouped per sheet.
• sheets[].layers — layer names, colorHex, and entityCount per sheet.
• file and summary — format, version, units, total entity count, and sheet count.
• imageUrl — a signed URL to the rendered PNG preview, ready to display.

Display the PNG in your UI and process the JSON in your backend — from one API call, with no separate rendering step and no additional infrastructure.