Peak Signal-to-Noise Ratio / Structural Similarity Index / Learned Perceptual Image Patch Similarity

Three widely used metrics for how similar a generated image is to a reference image, each closer to human perception than the last.

All three compare a generated or reconstructed image against a reference image, one pair at a time. PSNR (Peak Signal-to-Noise Ratio) converts mean squared error into decibels; typical values for 8-bit images fall between 30 and 50 dB, with higher being better, but it only looks at per-pixel differences and correlates poorly with what humans actually perceive. SSIM (Structural Similarity Index), published by Wang, Bovik, and colleagues in 2004, compares local image patches along luminance, contrast, and structure, with a maximum score of 1, higher is better. LPIPS, proposed by Zhang and colleagues at CVPR 2018, feeds both images through a pretrained deep network and compares their intermediate-layer features; lower is better, and it aligns with human perceptual judgment more closely than the other two. In embodied AI, all three are commonly used to evaluate novel-view synthesis, Gaussian-splatting reconstructions, and world-model predictions, but a high score does not guarantee the result is physically plausible.

ExampleThe original 3D Gaussian Splatting paper reports PSNR 27.21, SSIM 0.815, and LPIPS 0.214 on the Mip-NeRF360 dataset, matching the image quality of the then-best Mip-NeRF360 method.

Also called

PSNR, SSIM, LPIPS, Structural Similarity Index Measure

Related

Novel View Synthesis · 3D Gaussian Splatting · World Model · Fréchet Video Distance · Fréchet Inception Distance · Neural Radiance Fields

Sources

The Unreasonable Effectiveness of Deep Features as a Perceptual Metric (LPIPS, CVPR 2018)
Wikipedia: Structural similarity index measure
Wikipedia: Peak signal-to-noise ratio