An image generation model learns a latent visual space and samples from it to create new images, including images that interpolate between real examples. section 17.0
Diffusion models generate images by repeatedly denoising pure noise; conditioning that process on text produces images matching captions. section 17.4
Choose the model family according to the task:
- Use diffusion models for highly realistic generation and text-guided synthesis. section 17.4
- Use variational autoencoders for structured, continuous latent spaces supporting image editing and animation. section 17.4
- Treat GANs as an established alternative that has largely been displaced by diffusion models in recent practice. section 17.1.1
Training an unconditioned diffusion model
Train a diffusion model on flower images, then sample generated flowers after 100 epochs. Continue training to obtain increasingly realistic results; add text conditioning to move toward caption-controlled generation. section 17.2.7
model = DiffusionModel(image_size, widths=[32, 64, 96, 128], block_depth=2) model.normalizer.adapt(dataset) #1
Inspect the diffusion-model construction and dataset normalization used before training. section 17.2.7
