Can Random Noise Become a Pet? Train a Diffusion Model From Scratch in PyTorch | LIVE

Can a neural network learn to turn random static into a brand-new image? In this live build, we’ll train our own diffusion model and follow the whole journey—from real pet photos to noise, and from noise back toward a generated image. We’ll work with 64×64 color photos from the Oxford-IIIT Pet dataset and build a compact, multi-scale U-Net with about 770,000 trainable parameters. Starting with randomly initialized weights, we’ll teach it to predict the exact noise added to each image—then use that learned skill in a 300-step reverse diffusion process. In this session, we’ll: Watch real images gradually disappear into noise See how the model learns from noise it can verify for itself Train the U-Net with PyTorch, MSE loss and backpropagation Follow the full training loop, checkpoints and generated samples Run the Python scripts and guided notebooks, with the code explained as we go No pretrained weights are doing the learning for us. We’ll build and train the model ourselves. The first generations may be fuzzy; that gives us a chance to see what the model has learned and what more training would take. Join live, follow along, and watch the model try to turn static into something recognizable.