Is a higher LPIPS score better?

In the usual distance interpretation, lower indicates closer resemblance to the reference.

Can scores from different backbones be compared directly?

They should not be assumed comparable; use a consistent model and preprocessing for a controlled comparison.

Does LPIPS replace human evaluation?

No. It approximates aspects of perceptual judgment and should be checked against the actual task and relevant human assessments.

AI Glossary / Computer vision and multimodal AI / Intermediate

Learned perceptual image patch similarity (LPIPS)

Definition

Learned Perceptual Image Patch Similarity, usually abbreviated LPIPS, is a learned measure of visual difference between two images. It compares features extracted by a neural network rather than relying only on corresponding pixel values. In its usual interpretation, a lower score means that the images are more similar according to the chosen LPIPS model and evaluation setup.

What it means

Two images can have similar average pixel errors yet look very different to a person. A blurred image, a noisy image, and an image with a small spatial shift may affect perceived quality in different ways. LPIPS uses internal visual representations that respond to patterns beyond isolated pixel differences.

Despite the word similarity in its name, the commonly reported result behaves as a perceptual distance: smaller is closer. It is not a percentage of human agreement or a universal probability that two pictures look the same.

LPIPS is normally used with a reference image. It therefore answers a comparison question, not whether an image is attractive, truthful, or useful without a reference. A low score cannot by itself certify that a reconstruction preserves every important detail.

How it works

Pass both images through the same selected feature-extraction network. Compare normalized feature activations from several layers, apply learned channel weighting, and aggregate differences across spatial positions and layers.

The learned calibration is based on human perceptual comparisons. The original research studied how well different deep representations agree with judgments about distorted image patches. It did not establish that one score perfectly models perception in every application.

Prepare images using the implementation's required channel order, numerical range, dimensions, and preprocessing. The authors' common interface expects RGB tensors scaled to the appropriate range, so an unnoticed input-format error can invalidate comparisons.

Keep the backbone, learned weights, software version, and aggregation method fixed when comparing experiments. Different configurations can give different numerical scales and rankings; scores from incompatible setups should not be casually combined.

Three practical examples

1. Image restoration: A research team compares reconstructed photographs with their clean references, using LPIPS alongside pixel-based errors and visual inspection.

2. Compression evaluation: An imaging team checks whether candidate compression settings preserve visual appearance on a representative test collection, then compares the metric with human review.

3. Model training: A developer includes a differentiable perceptual comparison in a reconstruction objective, while separately checking content accuracy and other task requirements.

Think of it this way

Think of comparing two drawings by examining their shapes and textures rather than counting every differently colored square on a grid. The analogy is limited because LPIPS uses learned numerical features, not a person's understanding of the scene. It can overlook a meaningful change or react strongly to one a particular viewer would tolerate.

Worked example

Suppose two restoration systems are evaluated on three reference images using exactly the same LPIPS configuration. System A receives distances of 0.10, 0.20, and 0.30. Its mean is 0.60/3 = 0.20.

System B receives 0.15, 0.15, and 0.18, giving a mean of 0.48/3 = 0.16. On this tiny set, B has a lower average perceptual distance. The absolute mean difference is 0.04, or 20% of A's mean.

That relative reduction does not mean that people find B twenty percent better. LPIPS units are not a direct scale of user satisfaction, and three images provide little evidence about general performance. A is also better on the first image, so the average hides an individual tradeoff.

A fuller evaluation would examine paired differences across a larger test set, report uncertainty and difficult cases, and check whether the ranking agrees with the task's human or operational requirements. These invented scores illustrate interpretation only; they are not benchmark results.

Illustrative case study

Fictional case study: A document-imaging startup compares two restoration models for photographs of instruction labels. One produces sharper-looking texture and a lower average LPIPS score against the available reference images.

During review, the team discovers that this model occasionally changes a small printed digit. The global perceptual score does not adequately capture the consequence for the intended use. The team adds text-accuracy checks and targeted inspection of critical regions.

It retains LPIPS as one useful measure of overall visual resemblance, but chooses models using the combined evidence. The case illustrates a mismatch between a perceptual metric and a specific content requirement, without claiming a general failure rate for LPIPS or any real product.

Common misconceptions and mistakes

Lower LPIPS does not prove that an image is factually correct. Visually plausible textures can still contain wrong details.

LPIPS is not interchangeable with classification accuracy, aesthetic quality, or semantic similarity between unrelated scenes. Those are different questions.

A score should not be called a percentage unless a separate, justified transformation defines such an interpretation. A value of 0.2 does not automatically mean eighty percent similarity.

Using a metric as a training loss can improve that metric without guaranteeing improvement on every evaluation criterion. Also check preprocessing carefully: a channel-order or scaling error can create apparent model differences that come from the evaluation pipeline.

Business use

LPIPS can make evaluation of image-processing systems more informative when pixel error alone misses visible changes. It is particularly useful for comparing outputs against references under a consistent experimental setup.

For a business decision, use a representative image set and define what matters beyond appearance, such as readable text, preserved geometry, or acceptable artifacts. Record the metric configuration with each result and keep examples available for review. Treat the score as evidence that supports a decision, rather than a universal acceptance threshold shared by all types of imagery.

Check your understanding

A model's mean LPIPS falls from 0.25 to 0.20. Does this prove a twenty percent increase in perceived quality?

Answer: No. It is a twenty percent reduction in that mean distance, but the metric is not a direct percentage scale of perceived quality.

Related terms

Human evaluation
Perceptual loss
feature embedding
image quality assessment
PSNR
SSIM
reference image

Learn next

Compare pixel-based errors with feature-based distances on blur, noise, and spatial shifts. Then inspect where metric rankings disagree with human judgments or content-specific requirements, using an unchanged LPIPS configuration.

Sources

Zhang and colleagues, The Unreasonable Effectiveness of Deep Features as a Perceptual Metric: https://arxiv.org/abs/1801.03924

Authors' LPIPS implementation and usage guidance: https://github.com/richzhang/PerceptualSimilarity

The primary paper and implementation documentation were opened. The numerical comparison and document-imaging scenario are original illustrations.

Previous: Learned operator Next: Learned sparse retrieval