CGTuebingen/InterpTok-Models

InterpTok

Weights of Image Tokenizers Should Be Trained Between Images, by Ole Beißwenger, Jan-Niklas Dihlmann and Hendrik Lensch, University of Tübingen.

📄 Paper  |  💻 Code  |  🌐 Project Page

InterpTok fine-tunes the image tokenizer MAETok to decode mixtures of the codes of two images into natural images. On ImageNet at 256 × 256, it keeps MAETok's reconstruction quality and lowers its interpolation FID from 14.42 to 4.30. A SiT-B trained on its codes improves its generation FID without guidance from 11.61 to 8.48.

File	Model
interptok_41k.safetensors	InterpTok, the model of the paper, 41,000 steps
interptok_82k.safetensors	InterpTok trained further, to 82,000 steps
interptok_decoder_41k.safetensors	InterpTok's decoder on MAETok's frozen encoder, which decodes MAETok's codes and the samples of generators trained on them
sit-b_interptok-41k/	SiT-B trained for 400,000 steps on the codes of interptok_41k
sit-b_maetok/	SiT-B trained for 400,000 steps on MAETok's codes, the baseline of the paper

Each tokenizer holds 217M parameters in float32, MAETok-B-128's encoder and a decoder of 18 blocks, and encodes a 256 × 256 image into a code of 128 tokens with 32 values each. Each SiT-B folder is an experiment of the code of Xu et al.: args.json, the moving average of the weights in checkpoints/0400000.pt, and the configuration of its tokenizer. SHA256SUMS.txt lists the checksum of every file.

Usage

git clone --recursive https://github.com/cgtuebingen/InterpTok
hf download CGTuebingen/InterpTok-Models --local-dir InterpTok/weights
cd InterpTok
python src/interpolate.py a.jpg b.jpg --weights weights/interptok_41k.safetensors --out a_to_b.png

The code repository explains installation, training and evaluation.

Citation

@article{beisswenger2026interptok,
  title   = {Image Tokenizers Should Be Trained Between Images},
  author  = {Bei{\ss}wenger, Ole and Dihlmann, Jan-Niklas and Lensch, Hendrik},
  journal = {arXiv preprint arXiv:2610.16759},
  year    = {2026}
}

License

Apache License 2.0.