Deep learning

Image generation

product imagery, data augmentation, style transfer

Produce new images from noise, text, or other images.

Probability & Bayes

Maximum likelihood

Information theory

Optimisation

Foundations

VAEs maximise a likelihood bound; diffusion is a chain of small Gaussian steps; KL divergence appears in every loss.

Frame the problem

Frame

Unconditional, class-conditional, or text-conditional. Decide what quality means and who judges it.

Sourcing & signal

Labeling & data collection

Exploring the data

Data & labels

Captions if text-conditional. Check for duplicates, watermarks, and licensing before training on anything.

Modalities → vectors

Tokenization

Represent

Pixels, or a learned latent from an autoencoder. Text through a frozen encoder.

Split the data

Split

Hold out prompts and subjects, not just images.

Dumb baseline

Autoencoders

Convolutional networksbackground

Latent-variable models

VAEs for a smooth latent space and easy training; flows when you need exact likelihoods.

Variational autoencoders

Normalizing flows

Adversarial

Sharp samples, unstable training. Still used for speed.

GANs

Diffusion

The default today. Fine-tune a pretrained diffusion model rather than training from scratch.

Diffusion models

Vision transformers & CLIP

Model

An autoencoder is the baseline and the latent space the others build on.

Training craft

Fine-tuning a pretrained model

Fine-tuning & LoRA

Train

LoRA on a pretrained diffusion model covers most needs. From scratch only with a large, clean corpus and a budget.

Evaluating generative output

Evaluate offline

FID and CLIP score as proxies; human preference is the real metric. Check for memorised training images.

A/B testing & interleaving

Evaluate online

Preference tests against the current generator.

Inference optimisation

Serving & release

Ship & monitor

Sampling steps against latency. Distil or cache for production.

React Flow