July 8, 2025 · 1 min read
Notes on Diffusion Models
Working notes on forward/reverse diffusion and practical training tips.
generative-models
diffusion
deep-learning
Diffusion models have become the dominant approach for high-quality image generation.
Forward Process
def forward_diffusion(x_0, t, noise_schedule):
alpha_bar_t = noise_schedule.alpha_bar[t]
noise = torch.randn_like(x_0)
return torch.sqrt(alpha_bar_t) * x_0 + torch.sqrt(1 - alpha_bar_t) * noise, noiseClassifier-Free Guidance
eps = eps_uncond + guidance_scale * (eps_cond - eps_uncond)Training Tips
- Use EMA weights for sampling
- Cosine noise schedules outperform linear schedules
- Monitor FID but rely on human evaluation for final assessment