Read the original at HF Daily Papers
SALD is a self-referenced training framework for diffusion models that evaluates image-caption pairs at two noise levels using the same model to generate training signals without external demonstrations or teacher networks.
Carried by: HF Daily Papers. First seen: .