Song Yan

Organizes consistency in diffusion-based visual generation as three agreement relations—external, internal, and normative—and reviews representative methods, evaluation protocols, cross-category trade-offs, and open problems across image, video, and 3D generation.

Identifies information-gain collapse as a key obstacle to long latent visual reasoning and proposes SCOLAR, which generates auxiliary visual tokens from full-sequence hidden states while independently anchoring them to the original visual space. With staged supervised fine-tuning and ALPO reinforcement learning, it improves long-sequence latent reasoning, real-world benchmark performance, and out-of-distribution generalization.

Challenges the conventional view that initial noise in diffusion generation is merely random — reveals analyzable semantic patterns and proposes a training-free Semantic Erasure-Injection process.

Introduces StyleWallfacer — a unified framework for high-quality style transfer addressing semantic drift, overfitting, and color limitation through triple diffusion and semantic style injection.

Proposes SHSRD — an efficient conditional diffusion framework for hyperspectral image superresolution using spectral information injection and two-stage transfer learning on small HSI datasets.

Studies how a source diffusion model can be reused for domain perception and adapted through fine-tuning, aiming to improve few-shot image generation under limited target-domain data.

Explores dynamic knowledge for kernel modulation in one-shot multi-domain adaptation, targeting image generation when only extremely limited domain-specific supervision is available.

Formulates few-shot domain adaptation with dual objectives, jointly considering image restoration quality and cross-domain alignment to improve transfer under scarce target-domain examples.

Proposes DREAM, a training-free generative replay framework for class-incremental learning. It addresses the synthetic-real domain shortcut by using subspace rectification, orthogonal projection, and real-anchored prototype regularization to improve exemplar-free continual learning.