This repository contains the official PyTorch implementation for the paper:
Flow Map Distillation Without DataState-of-the-art flow models achieve remarkable quality but require slow, iterative sampling. FreeFlow explores a data-free alternative to flow map distillation, which conventionally requires external datasets. By sampling only from the prior distribution, our method circumvents the risk of Teacher-Data Mismatch. It learns to predict the teacher's sampling path while actively correcting for its own compounding errors, achieving state-of-the-art performance. Specifically, it reaches an impressive FID of 1.45 on ImageNet 256x256, and 1.49 on ImageNet 512x512, both with only 1 sampling step. We hope this work establishes a more robust paradigm for accelerating generative models.
We provide an
environment.yml file that can be used to create a Conda environment. If you only want to run pre-trained models locally on CPU, you can remove the
cudatoolkit and
pytorch-cuda requirements from the file.
1conda env create -f environment.yml
2conda activate DiT
Pre-trained FreeFlow checkpoints are hosted on the
Hugging Face organization page. You can sample from our pre-trained models with
sample.py. To use them, visit the Hugging Face download
guide, and pass the file path to the script, as shown below.
The script allows switching between the 256x256 and 512x512 models and changing the classifier-free guidance scale, etc. For example, to sample from our 512x512 FreeFlow-XL/2 model, you can use:
1@article{tong2025freeflow,
2 title={Flow Map Distillation Without Data},
3 author={Tong, Shangyuan and Ma, Nanye and Xie, Saining and Jaakkola, Tommi},
4 year={2025},
5 journal={arXiv preprint arXiv:2511.19428},
6}
