Li Fei-Fei

We propose grafting, a simple approach to materialize new architectures by editing pretrained diffusion transformers. It enables architectural exploration under small compute budgets.

Generating diverse plausible images with GFlowNets and diverse condition latent representation in conditional image generation