Diffusion Models From Eight Perspectives — DeepMind Scientist Analysis

Diffusion models power DALL-E 3, Stable Diffusion, and Sora. This analysis examines them through eight distinct lenses.

Perspectives

1. Probabilistic: Reverse a gradual noising process, optimizing a variational lower bound on data likelihood.
2. Autoencoder: Forward process encodes data to noise; reverse process decodes noise to data.
3. Energy-Based: Score matching learns gradients of the log-density toward high-probability regions.
4. SDE: Continuous-time formulation enables deterministic sampling via probability flow ODEs.
5. Markov Chain: Fixed-length chain (typically 1000 steps) with learned Gaussian transition kernels.
6. Score Matching: Model estimates the score function at each noise level.
7. Recurrent: Iterative denoising resembles an RNN with tied weights across timesteps.
8. Latent Variable: Diffusion in compressed latent space (Stable Diffusion) reduces computational cost.

Conclusion

These eight perspectives collectively explain diffusion model performance — combining score-matching theoretical rigor with practical engineering innovations.