perceptual loss

A loss or distance between images computed on the internal feature mapsfeature map, n.: The grid of numbers produced when one learned pattern detector is scanned across an image, showing where, and how strongly, that pattern appears. of a pretrained network (classically VGG) rather than on raw pixels. Pixel-wise losses such as mean squared errormean squared error, n.: A way of scoring numerical predictions: square each prediction's error, then average them, so big misses count much more than small ones. Lower is better. heavily penalize small misalignments and reward blurry averages; feature-space losses compare edges, textures and object parts instead, so networks trained with them produce sharper, more natural images. Perceptual losses are used in style transferstyle transfer, n.: Repainting a photo in the style of an artwork — your street as Van Gogh might have painted it — by keeping the photo's content and borrowing the painting's colors, strokes and textures., super-resolutionsuper-resolution, n.: Making a small or blurry image larger and sharper by having AI fill in plausible fine detail, rather than just stretching the pixels. and the training of the autoencoders behind VQGANVQGAN, n.: An image compressor that turns a picture into a small grid of codebook 'tiles' while keeping it sharp, used as the vocabulary of many image generators and of an early wave of AI art. and latent diffusionlatent diffusion model, n.: A diffusion image generator that works on a shrunken, compressed version of each picture instead of the full-size pixels, which makes it far cheaper to train and run.. LPIPS (learned perceptual image patch similarity) calibrates such feature distances against large sets of human similarity judgments and is widely used both as a training loss and as an evaluation metric; it agrees with human perception far better than traditional measures such as PSNR and SSIM.