LPIPS gives negative value

[mshooter]

Hi everyone,

my pseudo code is

loss = LearnedPerceptualImagePatchSimilarity(net_type="vgg", normalize=True) 
y_hat = model(x) 
loss(x, y_hat) 

both x and y_hat have values between [0,1].
But I don't understand why my loss is negative..

[F-Barto]

Same but with input values between [-1,1].
I am using it as a training loss.

[F-Barto]

Just opened an issue: #2315

[Borda]

@mshooter @F-Barto — this is a known issue and there are two common causes:

Cause 1: Input range mismatch. This is the most common one.

• normalize=False (default) expects inputs in [-1, 1]
• normalize=True expects inputs in [0, 1] and internally rescales to [-1, 1]

If your inputs are in [0, 1] but normalize=False, the network sees only half the expected range and can produce negative scores.

from torchmetrics.image.lpip import LearnedPerceptualImagePatchSimilarity

# If your images are [0, 1]:
lpips = LearnedPerceptualImagePatchSimilarity(net_type="vgg", normalize=True)

# If your images are [-1, 1]:
lpips = LearnedPerceptualImagePatchSimilarity(net_type="vgg", normalize=False)

Cause 2: Using LPIPS as a training loss. When backpropagating through the LPIPS network, the learned linear layers can yield negative values. The LPIPS paper states scores should be non-negative for detached evaluation, but during training gradient flow can push intermediate values negative.

Workaround if you need LPIPS as a loss:

loss = lpips(pred, target)
loss = torch.clamp(loss, min=0.0)  # safety clamp

Verify your setup:

import torch
img1 = torch.rand(4, 3, 64, 64)  # [0, 1] range
img2 = torch.rand(4, 3, 64, 64)

lpips = LearnedPerceptualImagePatchSimilarity(net_type="vgg", normalize=True)
score = lpips(img1, img2)
print(score)  # should be positive, typically 0.0–1.0

Docs: LPIPS