Diffusion & score-based generation

Entraînement

A network predicts noise, clean data, velocity, or a score at randomly sampled noise levels. Related flow-matching formulations learn a vector field between distributions.

Exécution de l’inférence

Sampling integrates a learned reverse process across multiple steps. Sampler choice, step count, guidance, and seed trade speed, diversity, and fidelity.

Atouts

• •Stable training relative to adversarial objectives
• •Flexible conditioning and editing
• •Strong coverage across image, audio, video, and scientific data

Compromis

• •Iterative sampling is slower than a one-pass generator
• •Guidance can reduce diversity or introduce artifacts
• •The data representation and sampler materially affect results

À utiliser lorsque

• High-dimensional conditional generation is central
• Multiple inference steps fit the latency budget
• Controls and sample quality can be evaluated jointly

À éviter ou remettre en question lorsque

• One-pass generation is mandatory
• The model must run on a severely constrained device
• “Diffusion” is being treated as one fixed implementation