Highly Compressed Tokenzier Can Generate Without Training

Tokenizers with highly compressed latent spaces -- such as TiTok, which compresses 256x256 px images into just 32 discrete tokens -- can be used to perform various image generative tasks without training a dedicated generative model at all. In particular, we show that simple test-time optimization of tokens according to arbitrary user-specified objective functions can be used for tasks such as text-guided editing or inpainting.

This repo includes the simple test-time optimization algorithm used in our ICML 2025 paper, "Highly Compressed Tokenzier Can Generate Without Training", under the tto/ directory.

For convenience, we include the TiTok implementation copied from the official code release under titok/.

Examples

• Text-guided image editing: notebooks/clip_opt.ipynb for test-time optimization with CLIP objective.

  

• Inpainting: notebooks/inpainting.ipynb for inpainting via reconstruction objective.

  

Running Locally: If you use Nix, you can enter a shell with all dependencies via nix develop.