LPIPSPipe

LPIPSPipe is registered as lpips-pipe and implemented in src/autopipeline/components/modules/lpips_pipe.py.

It is the framework's perceptual-distance module for full-image or region-limited comparisons.

Class

Overview

Registry Entry

Field	Value
Registry key	lpips-pipe
Class	LPIPSPipe
Main mixins	LPIPSMixin, MaskProcessor
Return type	float

Constructor

Constructor

LPIPSPipe(**kwargs)

Supported init kwargs

Key	Required	Default	Meaning
device	No	auto	Torch device for LPIPS inference.
net	No	alex	LPIPS backbone name passed to lpips.LPIPS.

Methods

Public Methods

Method	Purpose
calc_lpips(ref_image, edited_image, mask=None)	Compute LPIPS, optionally restricting the effective comparison region.
__call__(...)	Build the mask and dispatch to calc_lpips.

Signature

Call Signature

LPIPSPipe.__call__(
    ref_image: Image.Image,
    edited_image: Image.Image,
    coords: List[Tuple[int, int, int, int]] = None,
    mask_mode: str = None,
    metric: str = "lpips",
    **kwargs,
)

Input / Output

Runtime Inputs

Argument	Required	Meaning
ref_image	Yes	Reference image.
edited_image	Yes	Edited image.
coords	No	Pixel-space boxes to measure or exclude.
mask_mode	No	inner or outer, normally derived from pipeline scope.
metric	Yes	Currently only lpips.

Input / Output

Mask Behavior

LPIPSPipe does not crop the image before inference. Instead, when a mask is present it:

1. fills the unmeasured region of the edited image with pixels from the reference image
2. computes LPIPS on the composited edited image against the original reference image
3. scales the resulting score by the measured-area fraction

This area normalization is implemented as:

lpips_score / (computed_area / img_area)

when the mask contains at least one selected pixel.

Input / Output

Return Value

The pipe returns a single float:

• lower is better
• 0.0 means perceptually very similar
• larger values indicate stronger perceptual drift

Config

Minimal Config Example

metric_configs:
  lpips:
    pipe_name: lpips-pipe
    default_config: ${pipes_default.lpips-pipe}
    init_config:
      net: alex
    scope: unedit_area

Failure Mode

Failure Modes

This module is relatively stable because it has few branching paths.

The main preconditions are:

• coords must be valid if mask_mode is used
• input images must already be valid RGB-like arrays when converted through numpy

Unsupported metrics raise:

ValueError(f"Unsupported metric: {metric}")

Extension

Extension Notes

• Reuse this pipe if you still want a perceptual-distance metric with region support.
• Add a new pipe only if the metric semantics are no longer LPIPS-like.
• If you change area normalization, document the new score interpretation because downstream thresholding may depend on it.