F2D2

KO | EN

gitlite — search

#javascript #python #hacktoberfest #react #ai #android #rust #deep-learning #linux #docker #nodejs #windows #vue #openai

(★ 23)

Joint Distillation for Fast Likelihood Evaluation and Sampling in Flow-based Models

F2D2 최신버젼 다운로드

최종 버전 다운로드 (.zip)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# Joint Distillation for Fast Likelihood Evaluation and Sampling in Flow-based Models

The implementation for [Joint Distillation for Fast Likelihood Evaluation and Sampling in Flow-based Models](https://arxiv.org/abs/2512.02636).

![demo](https://raw.githubusercontent.com/Keely-Ai/F2D2/main/assets/comparison_grid.png)

Update: The Jax version of our implementation can be found [here](https://github.com/Keely-Ai/JAX-F2D2).

## Environment
Run the following command to set up the Python environment through `conda`.

```
conda env create -f f2d2_env.yml
```

## Datasets
Prepare the CIFAR10 dataset in the same way as [EDM](https://github.com/NVlabs/edm?tab=readme-ov-file#preparing-datasets).

## Training
1. Train a teacher Flow Matching model. Remember to set the variables in `run_fm.sh`.
```
run_fm.sh <NGPUs> <PORT>
```

2. Use the teacher Flow Matching model to train a Shortcut-Distill model. Remember to set the variables in `run_sc.sh`.
```
run_sc.sh <NGPUs> <PORT> 
```
3. Resume the Shortcut-Distill model to train a Shortcut-Distill-F2D2 model. Remember to set the variables in `run_ll.sh`.

```
run_ll.sh <NGPUs> <PORT> 
```

## Evaluation
Load checkpoints and evaluate the negative loglikelihood.
```
python nll.py \
  --test_batch_path <TEST_BATCH_PATH> \
  --model_pkl <MODEL_PATH> \
  --batch_size 128 \
  --time_eval_steps 8
```

## References
This repo is based on:
- [EDM](https://github.com/NVlabs/edm)
- [ECT](https://github.com/locuslab/ect)
- [Flow Matching](https://github.com/facebookresearch/flow_matching)
- [Consistency Model](https://github.com/openai/consistency_models)
- [CTM](https://github.com/sony/ctm)
- [Meanflow](https://github.com/Gsunshine/py-meanflow)

We would like to express our appreciation to them.

## Citation

```
@inproceedings{
    ai2026joint,
    title={Joint Distillation for Fast Likelihood Evaluation and Sampling in Flow-based Models},
    author={Xinyue Ai and Yutong He and Albert Gu and Ruslan Salakhutdinov and J Zico Kolter and Nicholas Matthew Boffi and Max Simchowitz},
    booktitle={The Fourteenth International Conference on Learning Representations},
    year={2026},
    url={https://openreview.net/forum?id=8uZ5UdIul2}
}
```

F2D2

F2D2 최신버젼 다운로드

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

F2D2

F2D2 최신버젼 다운로드

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

Download README