ByteDance's Seed team, in partnership with the University of Hong Kong, introduced DanceGRPO. This framework adapts the Group Relative Policy Optimization technique, originally from DeepSeekMath in April 2024, for visual generation tasks including text-to-image and video.
DanceGRPO modifies the sampling process to integrate group-based relative scoring into diffusion and rectified flow models. The framework supports models such as Stable Diffusion, FLUX, HunyuanVideo, and SkyReels-I2V, offering at least five distinct reward types.
Performance metrics show significant gains, with DanceGRPO achieving improvements of up to 181% on HPS-v2.1 and CLIP Score benchmarks. A subsequent project, BranchGRPO, announced in September 2025, increased alignment scores by 16% while reducing training time by 55%.
Additionally, ByteDance collaborated with Tsinghua University's AIR lab to develop DAPO for language model reasoning. This method reached a score of 50 on the AIME 2024 benchmarks, utilizing 50% fewer training steps compared to a DeepSeek-R1 configuration.
