Recently, the MixGRPO team from Tencent's Yuan Foundation Model released a breakthrough image generation framework called MixGRPO. This approach not only shortens the training time by nearly 50%, but also performs excellently, introducing a variant called MixGRPO-Flash, which further reduces the training time by 71%. All of this is made possible by their innovative sampling strategy that combines stochastic differential equations (SDE) and ordinary differential equations (ODE).
In current image generation technologies, efficiency and quality often conflict. MixGRPO improves the Markov decision process (MDP) by introducing a hybrid sampling method, significantly enhancing training efficiency. Specifically, the framework limits the random exploration range of agents, reducing computational costs during the optimization process while simplifying the model training process.
Compared to the previous DanceGRPO model, MixGRPO shows significant improvements in multiple dimensions. The research team demonstrated through experiments that optimizing specific denoising steps is sufficient to maintain or even enhance performance. The study also points out that although MixGRPO reduces training time and computational costs, it also requires the introduction of higher-order solvers to accelerate the sampling of old policy models.
In addition, MixGRPO adopts a sliding window strategy, allowing the model to gradually focus on more critical time steps during the denoising process, achieving more efficient optimization. This innovation has brought considerable progress in the diversity and quality of image generation.
MixGRPO not only opens up new directions for the future of image generation technology, but also provides valuable experience and reference for subsequent research. The open-source code is provided at the end of the article, and we look forward to more developers joining this exciting technological exploration.
Project URL: https://tulvgengenr.github.io/MixGRPO-Project-Page/