Choosing a Tokenizer

medtokenizers provides several tokenizer families, all subclasses of BaseTokenizer with a common interface.

Continuous vs. discrete

• ContinuousTokenizer – VAE/AE-style models that map a scan to a continuous latent grid. Pair these with continuous generative models (diffusion, flow matching).

• DiscreteTokenizer – VQ / FSQ / LFQ / ResidualFSQ models that map a scan to a grid of integer codes. Pair these with discrete generators (autoregressive, MaskGIT, discrete diffusion) in medlatents.

Specialized tokenizers

• MAISITokenizer – a continuous tokenizer configured to the NVIDIA MAISI VAE layout, with sliding-window reconstruct for large 3D volumes.

• TiTokTokenizer – a transformer tokenizer that encodes an image or volume into a fixed-length 1D sequence of K tokens.

• RAETokenizer (experimental) – a Representation Autoencoder pairing a frozen foundation-model encoder with a trainable patch decoder.

Rule of thumb

Choose a discrete tokenizer if you intend to train discrete generators; choose a continuous tokenizer for continuous diffusion / flow models. TiTok is useful when a compact 1D sequence is preferable to a spatial token grid.