Diffusion Generative Models
Key Concepts
- Forward (q): adds noise deterministically. Reverse (p_θ): denoises via learned network
- ᾱₜ vs αₜ: αₜ = 1 − βₜ (single-step). ᾱₜ = ∏αₜ (cumulative). Forward closed form uses ᾱₜ
- Cosine schedule (Nichol & Dhariwal 2021): s=0.008 offset prevents ᾱ changing too fast near t=0
- DDIM η: η=0 → deterministic ODE sampler; η=1 → stochastic DDPM sampler
Score Matching
Learn the score ∇_x log p(x) instead of p(x) directly. Denoisin
[Description truncada. Veja o README completo no GitHub.]