5 ms·
A novel and memory-friendly preference alignment method for diffusion models that does not depend on any reference model.
by kashifr 2y ago
A novel and memory-friendly preference alignment method for diffusion models that does not depend on any reference model.