Representation Alignment
Concept
A method published in 2024 that couples the model's loss with a loss reflecting similarity between internal diffusion transformer representations and pre-trained encoder representations, significantly speeding up training.
Mentioned in 1 video
