Diffusion Model Fundamentals
Generative models via diffusion process.
How They Work
Forward process: add noise gradually. Reverse process: learn to denoise. DDPM: denoising diffusion probabilistic models.
Architecture
U-Net for image denoising. Conditioning via cross-attention. Classifier-free guidance.
Stable Diffusion
Latent diffusion: compress to latent space. Text conditioning: CLIP text encoder. Open weights, runs locally.
Key Takeaways
- Diffusion: add noise, learn to reverse
- Stable Diffusion: latent diffusion
- High-quality image generation