Command Palette
Search for a command to run...
LPIPS / Temporal LPIPS
LPIPS (Learned Perceptual Image Patch Similarity) was proposed by Richard Zhang et al. at CVPR in 2018 and is a deep-feature metric for measuring perceptual similarity between images. It uses a pretrained network (such as AlexNet or VGG) to extract multi-layer features and computes a weighted distance in feature space, which is closer to human subjective perception than PSNR and SSIM. A lower value indicates that two images are more perceptually similar, so the lower the value, the better. It is widely used in tasks such as image generation, super-resolution, image restoration, and editing to evaluate perceptual defects such as texture details and artifacts.
Here, Temporal specifically refers to Temporal LPIPS (LPIPS-based inter-frame perceptual distance), where lower is better; if the literature uses a positive Temporal Consistency score, the direction is opposite.
LPIPS computes frame-by-frame perceptual distance based on deep features to evaluate frame-level visual quality of videos; Temporal LPIPS computes the perceptual distance between adjacent frames to measure temporal consistency (a lower value indicates greater inter-frame stability). Both are perceptual distances, where lower is better. If a positive consistency-rate metric is used, the direction is opposite.
Build AI with AI
From idea to launch — accelerate your AI development with free AI co-coding, out-of-the-box environment and best price of GPUs.