May, 2024
超越构成式推理,DDPMs 能够产生零样本插值
Going beyond compositional generalization, DDPMs can produce zero-shot interpolation
Justin Deschenaux, Igor Krawczuk, Grigorios Chrysos, Volkan Cevher
TL;DRDenoising Diffusion Probabilistic Models (DDPMs) can effectively generate images in unexplored regions of the data distribution by composing latent factors learned from separate subsets, demonstrated through zero-shot interpolation for attributes like smiling faces.