# WAFT
WAFT, Warping-Alone Field Transforms, is an optical flow method by Wang and Deng (ICLR 2026). It replaces RAFT's cost volume with high-resolution feature warping and a ViT refinement loop. LibreYOLO ports the upstream code and mirrors its four zero-shot checkpoints, one per feature encoder.
Tasks: Optical flow. Install: pip install libreyolo.
Verified against LibreYOLO v1.8.0.

WAFT drops the correlation volume that RAFT and its descendants
build, and instead warps high-resolution features of the second frame with the
current flow estimate before each refinement step. All four checkpoints are
upstream's zero-shot models (TartanAir, then Chairs, then Things). Pick `m`,
the a1 checkpoint the upstream README recommends for downstream use, unless
you have a reason to choose a specific encoder.

## Install

WAFT needs no optional extra. Everything it imports is in the base install.

```bash
pip install libreyolo
```

## Predict

Weights download from the LibreYOLO Hugging Face organization on first use.
The class `LibreWAFT` is also exported from `libreyolo` if you prefer to build the
model directly.

**Python**

```python
from libreyolo import LibreYOLO

model = LibreYOLO("LibreWAFTm-flow.pt")   # weights download on first use
results = model(["frame_0001.jpg", "frame_0002.jpg"])

flow = results[0].flow
print(flow.data.shape)        # (H, W, 2) float32, (dx, dy) in pixels
print(flow.mean_magnitude)    # mean displacement in pixels
```

**Video**

```python
from libreyolo import LibreYOLO

model = LibreYOLO("LibreWAFTm-flow.pt")

# N frames give N - 1 results; result i is the flow from frame i to i + 1.
for result in model("video.mp4", stream=True, save=True, save_flow=True):
    print(result.flow.max_magnitude)
```

**CLI**

```bash
libreyolo predict model=LibreWAFTm-flow.pt source=frame_0001.jpg,frame_0002.jpg save=true
libreyolo predict model=LibreWAFTm-flow.pt source=video.mp4 save=true arrows=true save_flow=true
```

Flow needs at least two frames: a single image raises. A list of images, a
directory sorted by name, a video or a stream is one sequence, and `N` frames
give `N - 1` results. `result.flow` is a `FlowField` with an `(H, W, 2)` float32
array on the original canvas of the first frame; `flow.data[y, x]` is
`(dx, dy)` in pixels, `+x` right and `+y` down. WAFT fills no confidence map, so `flow.confidence` is `None`. `result.boxes` is
`None`. `save=True` writes the Middlebury colour wheel, `arrows=True` or
`warp=True` pick the other views, and `save_flow=True` also writes each pair's
raw field as a `.flo` file. The [optical flow task](/docs/tasks/optical-flow)
page describes the payload and views in full.

## Variants

| Size | Upstream checkpoint | Parameters | File |
| --- | --- | ---: | --- |
| `t` | a2, DINOv3 ViT-S/16 encoder | 56.5 M | `LibreWAFTt-flow-dinov3.pt` |
| `s` | a2, Depth Anything V2 ViT-S | 56.9 M | `LibreWAFTs-flow-dav2.pt` |
| `m` | a1, Depth Anything V2 ViT-S | 64.2 M | `LibreWAFTm-flow.pt` |
| `l` | a2, Twins-SVT-Large encoder | 136.0 M | `LibreWAFTl-flow-twins.pt` |

The encoder is the weight variant in the file name, and each size exists only
under the name in the table; any other combination raises with the published
name. `LibreWAFT` defaults to size `m`. Prediction runs at native resolution:
the network pads the pair to a multiple of 112, 32 or 16 depending on the
encoder and crops the flow back. Five refinement iterations are baked in, the
upstream evaluation default. `imgsz` caps the long side (default 8192, in
practice no cap).

Training is not offered: `train()` raises `NotImplementedError`.

## Validate

`val()` runs the shared flow validator on a flow dataset (frames per sequence
folder, Middlebury `.flo`, KITTI 16-bit PNG or `.npy` ground truth; see the
[dataset format](/docs/tasks/optical-flow#dataset-format)). It reports the mean
end-point error `metrics/epe`, `metrics/px1` and `metrics/px3` (share of valid
pixels under 1 and 3 px of error) and the KITTI outlier rate
`metrics/fl_all`.

**Python**

```python
from libreyolo import LibreYOLO

model = LibreYOLO("LibreWAFTm-flow.pt")
metrics = model.val(data="my_flow.yaml")

print(metrics["metrics/epe"])      # mean end-point error, pixels
print(metrics["metrics/px1"])      # share of pixels under 1 px error
print(metrics["metrics/fl_all"])   # KITTI outlier rate
```

**CLI**

```bash
libreyolo val model=LibreWAFTm-flow.pt data=my_flow.yaml
```

## Export

| Task | ONNX | TorchScript | ExecuTorch | TensorRT | OpenVINO | Paddle | MNN | RKNN | Axelera | ncnn | TFLite | CoreML | Core AI |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
| Optical flow |  |  |  |  |  |  |  |  |  |  |  |  |  |

A "yes" means the export is supported. An empty cell means the exporter refuses that combination.

Export is not available yet for any format. Optical flow export is blocked
until the two-input graph (`image1`, `image2`) and the backend vector-rescale
contract of the flow task are implemented for a family.

## Checkpoints

Every published weight file for this family.

| File | Input (px) | Task | Weights license |
| --- | --- | --- | --- |
| `LibreWAFTm-flow.pt` |  | Optical flow | bsd-3-clause |
| `LibreWAFTt-flow-dinov3.pt` |  | Optical flow | DINOv3 License |
| `LibreWAFTs-flow-dav2.pt` |  | Optical flow | bsd-3-clause |
| `LibreWAFTl-flow-twins.pt` |  | Optical flow | bsd-3-clause |

## Licensing

Check the license on the Hugging Face repository of the specific weights you download. That repository is authoritative and licenses are not always uniform across a family. This is a description of the licenses involved, not legal advice.

- Original work: WAFT, Princeton Vision & Learning Lab
- Upstream license: BSD-3-Clause (s, m and l checkpoints); DINOv3 License (t checkpoint)
- Upstream source: https://github.com/princeton-vl/WAFT
- LibreYOLO code: BSD-3-Clause
- Weights: BSD-3-Clause (s, m and l checkpoints); DINOv3 License (t checkpoint), republished at https://huggingface.co/LibreYOLO
- Interpretation: The license depends on the size, because each checkpoint embeds a different encoder. The s, m and l checkpoints are BSD-3-Clause, a permissive license that allows commercial and closed-source use as long as the copyright notice, conditions and disclaimer are kept on redistribution; their encoders are Apache-2.0 (Depth Anything V2 Small for s and m, Twins-SVT-Large from timm for l), which is also permissive. The t checkpoint contains a DINOv3 ViT-S/16 fine-tuned by the WAFT authors, so Meta's DINOv3 License applies to it: a custom license, not OSI-approved, with its own use restrictions and conditions; read it before relying on t commercially. Upstream publishes no separate weight license, so LibreYOLO hosts the checkpoints under the repository's BSD-3-Clause license plus the encoder license. The code from princeton-vl/WAFT at commit b152ff1 is vendored under libreyolo/models/waft/_vendor/ (BSD-3-Clause); the encoders reuse code already in LibreYOLO (Depth Anything V2, Apache-2.0; DINOv3 from the DEIMv2 vendor, DINOv3 License) and a plain-torch Twins copy derived from timm (Apache-2.0). LibreYOLO's wrapper code is MIT. Weights come from the upstream Google Drive model zoo and are mirrored on the LibreYOLO Hugging Face org without gating. Inference only.

WAFT is the work of Wang and Deng at the
[Princeton Vision & Learning Lab](https://github.com/princeton-vl/WAFT). The
upstream code is vendored under its BSD-3-Clause licence. Upstream publishes no
separate weight licence, so LibreYOLO mirrors the checkpoints under the
repository's BSD-3-Clause licence plus the licence of the encoder each one
embeds, without gating.

- `s`, `m` and `l` are BSD-3-Clause. Their encoders are Apache-2.0 (Depth
  Anything V2 Small for `s` and `m`, a Twins-SVT-Large copy derived from timm
  for `l`). Commercial and closed-source use is allowed.
- `t` contains a DINOv3 ViT-S/16 fine-tuned by the WAFT authors, so Meta's
  **DINOv3 License** applies to it: a custom licence, not OSI-approved, with
  its own use restrictions and trade-control terms. Read it before using `t`
  commercially.

## Citation

```bibtex
@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}
}
```

Copied from https://github.com/princeton-vl/WAFT
