MixGRPO: A New Open-Source Solution from Mixue with Significantly Improved Training Efficiency and Enhanced Performance
Tencent's MixGRPO framework combines SDE/ODE sampling to cut training time by 50%, with MixGRPO-Flash reducing it by 71%. It optimizes MDP via exploration constraints and sliding window denoising, enhancing image quality/diversity. Code is open-sourced.....