Repository navigation
JLD and JLD-fast are now available as an installable, versioned Python distribution with author-maintained PyIQA and TorchMetrics integrations. The original scoring operations and fitted lens are unchanged.
Install in a uv project:
uv add "jacobian-lens-distance @ https://github.com/shreshthsaini/jld/releases/download/v1.1.0/jacobian_lens_distance-1.1.0-py3-none-any.whl"Optional integrations are installed through the pyiqa and torchmetrics extras. The README documents PyIQA registration and the TorchMetrics update/compute/reset interface. Python imports remain jld.
- Paper: https://arxiv.org/abs/2610.05967
- HF Papers: https://huggingface.co/papers/2610.05967
- Fitted lens: https://huggingface.co/shreshthsaini/JLD
- Demo: https://huggingface.co/spaces/shreshthsaini/JLD-demo
Validation: 73 tests passed and one CUDA test skipped on CPU, plus eight demo tests. A clean noneditable uv installation verified both integrations, release-example score agreement, plugin entry points, and installed CLI/map output. Distributed Gloo tests verify sample weighting with unequal rank counts. No new benchmark result is claimed.
The wheel bundles the fitted lens; timm downloads the frozen encoder on first use. SHA256SUMS covers both distributions. PyPI publication uses the prepared Trusted Publishing workflow and will be verified separately.