Skip to content

LPIPS-SqueezeNet (family page)

LPIPS-SqueezeNet is a full-reference perceptual distance: the lpips feature extractor scores how different a distorted frame looks from its reference, using features of a pretrained SqueezeNet. This page explains the family and points at the versioned card; the usage, checkpoint facts and limitations live on the card.

Which card to read

Registry id Display name File Card
lpips_sq_v1 vmaf_tiny_lpips_sq_v1 model/tiny/lpips_sq.onnx lpips_sq_v1.md

The ONNX file is named lpips_sq.onnx (no version suffix), while the registry id and the card carry the version. Future checkpoints get their own versioned card next to lpips_sq_v1.md.

At a glance

  • Extractor: lpips, one value per frame pair; 0.0 means perceptually identical, larger means more different. The score is a ranking signal, not MOS-calibrated.
  • Run it: vmaf ... --feature lpips=model_path=model/tiny/lpips_sq.onnx (or set VMAF_LPIPS_MODEL_PATH). The full CLI, C API and Python usage are on the lpips_sq_v1 card.
  • Upstream: richzhang/PerceptualSimilarity v0.1 (SqueezeNet linear weights, BSD-2-Clause), exported by ai/lpips_export.py.
  • Size: 3.2 MB, ONNX opset 18, two inputs (ref, dist), scalar output score.

See also

  • overview.md — where LPIPS fits in the C1–C4 capability map
  • inference.md — loading and using tiny models from libvmaf
  • security.md — ONNX op-allowlist and registry sha256 pinning
  • ADR-0040 and ADR-0041 — multi-input session API and the extractor design