anycalib-mlx

AnyCalib — MLX weights

Converted weights for mlx-swift-AnyCalib, a Swift/MLX port of AnyCalib (Tirado-Garín & Civera, ICCV 2025): single-view camera calibration with a camera model chosen after the network runs.

These are not new weights. They are the official AnyCalib checkpoints in a different container.

What was changed

• Repacked from PyTorch .pt into safetensors.
• Convolution tensors transposed (out, in, kH, kW) → (out, kH, kW, in) for MLX's native NHWC layout. That same rule also reshapes the (1, 3, 1, 1) ImageNet mean/std buffers to (1, 1, 1, 3).
• Optionally cast to float16; the default build is float32.

No weight values are altered beyond that optional cast. The port is verified stage by stage against the PyTorch reference — on MLX's CPU backend every stage matches to within float32's own noise floor (ray field within 0.0003°).

Layout

One subdirectory per pretrained variant, each holding config.json and weights.safetensors:

anycalib_dist/
anycalib_edit/
anycalib_gen/
anycalib_pinhole/

The variants differ only in training imagery: pinhole (perspective only), gen (perspective + distorted), dist (distorted + strongly distorted), edit (stretched and cropped perspective).

Usage

anycalib calibrate photo.jpg --cam pinhole --repo mnmly/anycalib-mlx

The Swift loader resolves these through the shared Hugging Face cache, so a copy pulled by huggingface_hub is reused and vice versa.

License and attribution

Apache 2.0, inherited from upstream. See LICENSE and NOTICE in this repo.

Original work © Javier Tirado-Garín and Javier Civera, I3A, University of Zaragoza. The backbone is DINOv2, © Meta Platforms, Inc., also Apache 2.0.

@InProceedings{tirado2025anycalib,
  author={Javier Tirado-Gar{\'i}n and Javier Civera},
  title={{AnyCalib: On-Manifold Learning for Model-Agnostic Single-View Camera Calibration}},
  booktitle={ICCV},
  year={2025}
}