Projects
MultiMaxPO
Official PyTorch implementation of Multi-Axis Max@K Reinforcement Learning for representative diversity in text-to-image generation. The repository includes the MultiMaxPO advantage estimator, a synthetic demonstration, and training and evaluation code for SD3.5-M experiments.
AutoRef
Official implementation of AutoRef, a method that optimizes the executable harness around frozen image generation and reasoning models for agentic multi-reference image generation.