PerceptualSimilarity

Why we included this project

LPIPS scores how alike two images look to a person by comparing deep feature maps, which usually tracks human judgment far better than pixel-level metrics like PSNR or SSIM. That makes it a common choice for anyone training or evaluating image generation, super-resolution, or restoration models, since you can backpropagate through it as a perceptual loss. Setup is a single pip install and the call itself is just a distance between two normalized tensors, with pretrained AlexNet, VGG, or SqueezeNet networks doing the heavy lifting. The repo also ships the BAPPS dataset, handy when you want to verify that some other similarity metric actually agrees with human perception.