Turn images into structured JSON with custom fields

LFM2.5-VL-1.6B-Extract is a compact vision model built for structured image extraction. Instead of asking a general vision model to describe an image in paragraphs, you give it a field list and it returns only the requested values as JSON.

The practical workflow is simple. You write a small schema, such as `brand`, `product_type`, `net_weight`, or `expiry_date`, then pass in a product photo, receipt, nutrition label, business card, or safety image. The model reads the image and fills those fields directly, which makes the output easier to pass into a rule-based system, spreadsheet import, or downstream app.

The public Space makes the workflow more approachable because it includes a visual schema builder. You can add fields one by one, choose the 450M or 1.6B model, upload an image, and compare the returned card view with the raw JSON.

Start with the Liquid Image to JSON Space. Pick one of the presets, such as receipt, nutrition label, business card, product photo, or wood surface, then replace the example image with one of your own. For a first real test, use a receipt or label with 4-6 fields and check whether the model returns parseable JSON with the exact keys you asked for.

For local use, the model card includes a Transformers example. Install `transformers` and `pillow`, load `LiquidAI/LFM2.5-VL-1.6B-Extract` with `trust_remote_code=True`, and pass a system prompt that lists your fields in YAML. Keep the first local run single-image and single-turn, and use greedy decoding so the model stays focused on returning JSON.

This is a focused extraction model, not a general image reasoning assistant. The model card frames it around one image, one schema, and a flat JSON object, so test it on the exact field shape you need before trusting it in an automated workflow.

• Extract product attributes from catalog photos before importing them into a database.
• Pull totals, dates, tax fields, or payment details from receipts.
• Turn business cards into structured contact records.
• Flag visual safety events, such as leaks or fire, as fields an automation can read.
• Compare a small 450M model against the 1.6B model for speed and accuracy.

Try the demo

View model page