Why it matters

SD math is foundation for open image gen. Understanding shapes design.

Advertisement

The architecture

Forward: q(z_t | z_0) = N(sqrt(alpha_t) z_0, (1-alpha_t) I).

Reverse: predict noise.

Stable Diffusion mathForwardadd noiseUNetpredict noiseReversedenoise stepsLatent space (VAE-encoded) shrinks image to work in; 4x compressed
SD math.
Advertisement

How it works end to end

Loss: MSE between predicted + true noise.

Sampling: DDIM, DPM++, Euler.

Guidance for text conditioning.