Collected sources and patterns will appear here. Add from search or the patterns library.
NoisyLatent + ClassEmbedding + Timestep → DenoisedLatent (via patchified transformer + adaLN-Zero)
Replaces the U-Net backbone of latent diffusion with a pure transformer operating on patchified latent tokens; class or text conditioning is injected via adaptive layer norm zero-initialization (adaLN-Zero) which scales and shifts each token's representation conditioned on the timestep and class embedding, enabling scalable training through standard transformer scaling laws.
Problem it solves
U-Net's inductive biases limit scaling predictability; a pure transformer backbone obeys transformer scaling laws and enables architecture reuse across vision tasks.
Consumes
Emits
Establishes
The real projects this mechanism was found in. Attribution is the point — this is how the best teams actually do it.