偏向基础理论：

Representation Learning：

Revisiting Multi-Task Visual Representation Learning

DDRD 和Brownian Bridge 的区别

DDRD方案：

Decoupled Residual Denoising Diffusion Models for Unified and Data Efficient Image-to-Image Translation

Brownian Bridge Matching方案

DDBM:Denoising Diffusion Bridge Model

LBM: Latent Bridge Matching for Fast Image-to-Image Translation

Vision Bridge Transformer at Scale

Marigold:Repurposing Diffusion-Based Image Generators for Monocular Depth Estimation

Dense Prediction（分割 | 法向量 | 深度图）

分割

X-SAM: From Segment Anything to Any Segmentation

SemFlow: Binding Semantic Segmentation and Image Synthesis via Rectified Flow

深度|法相

What Matters When Repurposing Diffusion Models for General Dense Perception Tasks?

（使用vae latent 替代noise）E2E-FT: Fine-Tuning Image-Conditional Diffusion Models is Easier than You Think

（固定noise）FE2E: From Editor to Dense Geometry Estimator (cvpr2026)