Model

We offer two types of segmentation models for creating bounding boxes - Fast and HighQuality. This section details the differences between each model, and when a specific model should be used.

​

Configuration

You can select which model to use by setting the model parameter in the POST request body.

type Model = "Fast" | "HighQuality";

​

Fast

The Fast model uses LightGBM and runs on CPUs - which makes it much faster than the HighQuality model. It depends on a pre-existing text-layer to create bounding boxes.

This makes it great for text, table and formula segmentation - but it struggles when dealing with images. It can still segment images that are surrounded/bounded by text to some degree. A good example is graphs that have Y/X axes with labels.

The fast model can’t be used for PDFs without a text-layer.

​

HighQuality

The HighQuality model uses VGT - a state-of-the-art two-stream Vision Grid Transformer. This model ‘has eyes’ and can understand layouts with intelligence, so a pre-existing text-layer is not required. It’s better at classifying content into segment_type.

​

What will work best for me?

In most cases, we recommend using the Fast model for it’s speed and cost benefits, and using the HighQuality model as a fallback for when the content field in the output object is empty (i.e when the Fast model returns an empty array).

If you require great table, image or handwriting segmentation consistently, then you should use the HighQuality model.