CTM introduces a unified framework for generative models, improving sampling efficiency and quality by integrating score-based and distillation strategies.
Enhanced GAN-based denoising method SPI-GAN reduces sampling time while maintaining high quality and diversity.
Generalized Consistency Trajectory Models (GCTMs) extend CTMs to enable translation between arbitrary distributions, enhancing diffusion-based image manipulation tasks.
SPI-GANは、高いサンプリング品質と多様性を実現しつつ、サンプリング時間を劇的に短縮する手法であり、CIFAR-10とCelebA-HQ-256のデータセットにおいて最もバランスの取れたモデルの1つである。
提案されたDDMIは、高品質な暗黙のニューラル表現を生成するためのドメインに依存しない潜在拡散モデルです。
Geometric generative models based on morphological equivariant PDEs and GANs outperform traditional CNNs in image generation tasks.
Discrete Distribution Networks (DDN) offer a new approach to generative modeling by approximating data distributions using hierarchical discrete distributions, enabling efficient data representation and unique zero-shot conditional generation capabilities.