Lotus: Diffusion-Based Visual Foundation Model for High-Quality Dense Prediction