ENHANCE! Upscaling Images with Neural Networks
summary
Claims that modern GANs can plausibly upscale low-res photos into sharp outputs, using a convolutional generator and discriminator trained on 200,000 celebrity faces and paired low/high examples. Replacing simple per-pixel loss vs adversarial loss produces much sharper images, but the system hallucinates details it never saw. Built results look realistic yet often mismatch the true high-res source, fail on nonface logos, sometimes produce grotesque artifacts, and therefore work visually but are unsafe for forensic identification.
AI-written summary · about
chapters (5)
- 0:00Start
- 3:10Neural network upscaling
- 3:23Neural network training
- 6:47Adversarial games
- 9:01GAN-trained generator
related talks
43mHitchhiker’s Guide to Generative Adversarial Networks (GANs)
37mHarnessing the power of Generative Adversarial Networks (GANS) for supervised learning
37mLearning From Machines
44mLearning from Machines
40mOverview of Generative Adversarial Networks (GANs)
33mLow to High Resolution
34mLearning from machines
18mDance of the Ancestors...