Model types

A model is what a processing job runs with. Every job that recognises, detects or extracts takes a model ID. The model families below are what those IDs refer to.

Text recognition models

Two families perform text recognition, and which you pick matters more than any parameter.

Super models are large general-purpose models that work across many hands, languages and centuries without any training. Text Titan II is the current one. Use these when you have mixed or unfamiliar material, or when you have no ground truth. They are the right default.

PyLaia models are small, fast models trained on a specific hand, collection or script. They decode in a single pass, which makes them considerably cheaper at volume, and a well-trained one will beat a general model on the material it was trained for. Use these when you have a large homogeneous collection and either an existing public model that fits it or enough ground truth to train your own.

Both are addressed the same way, by model ID (htrId in the job config). See Browse models for where model IDs come from today.

Layout models

Layout models produce baselines, regions and reading order. The API currently exposes them only inside text recognition jobs, through the config.lineDetection parameters; the field name is historical, the models are layout models. See Layout.

Other model families

Field models (template-bound region detection) and table models (cell and structure detection) power Fields and Tables in Transkribus today. They are addressed by model ID like every other model.

Training your own model works in Transkribus today. See Model training.