[Paper]

We introduce Warping-Alone Field Transforms (WAFT), a simple and effective method for optical flow. WAFT is similar to RAFT but replaces cost volume with high-resolution warping, achieving better accuracy with lower memory cost. This design challenges the conventional wisdom that constructing cost volumes is necessary for strong performance. WAFT is a simple and flexible meta-architecture with minimal inductive biases and reliance on custom designs. Compared with existing methods, WAFT ranks 1st on Spring, Sintel, and KITTI benchmarks, achieves the best zero-shot generalization on KITTI, while being up to 4.1× faster than methods with similar performance.

If you find WAFT useful for your work, please consider citing our academic paper:

Yihan Wang, Jia Deng

@article{wang2025waft,
  title={WAFT: Warping-Alone Field Transforms for Optical Flow},
  author={Wang, Yihan and Deng, Jia},
  journal={arXiv preprint arXiv:2506.21526},
  year={2025}
}

• [2025/09] WAFT now achieves sota performance on all benchmarks and supports multiple backbones (Twins/DAv2/DINOv3).

Our code is developed with pytorch 2.7.0, CUDA 12.8 and python 3.12.

conda create --name waft python=3.12
conda activate waft
pip install -r requirements.txt

Please also install xformers following instructions.

Google Drive: link.

To evaluate/train WAFT, you will need to download the required datasets: FlyingChairs, FlyingThings3D, Sintel, KITTI, HD1K, TartanAir, and Spring. Please also check SEA-RAFT for more details.

You may need to prepare pre-trained DepthAnythingV2 or DINOv3 before you start training corresponding models.

python train.py --cfg config/a2/twins/chairs.json
python train.py --cfg config/a2/twins/chairs-things.json --restore_ckpt ckpts/twins/chairs.pth

python evaluate --cfg config/a2/twins/chairs-things.json --ckpt ckpts/twins/zero-shot.pth --dataset sintel
python submission --cfg config/a2/twins/tar-c-t-kitti.json --ckpt ckpts/twins/kitti.pth --dataset kitti

This project relies on code from existing repositories: RAFT, DPT, Flowformer, ptlflow, DepthAnythingV2, and DINOv3. We thank the original authors for their excellent work.