Researchers have introduced LC-GRPO, a novel framework for flow-based GRPO that incorporates Langevin correction to bridge the gap between training and inference in generative models. This method addresses the discrepancy that arises when using stochastic differential equations (SDEs) for training and deterministic ordinary differential equations (ODEs) for inference, which can lead to blurry samples and reduced performance. By applying a Langevin correction after an ODE Euler step, LC-GRPO theoretically reduces Wasserstein error and improves sample accuracy compared to standard SDE discretizations. Experiments on models like SD3.5-Medium, FLUX.1-Dev, and HunyuanVideo show consistent improvements in reward optimization for text-to-image and text-to-video generation while maintaining generation quality. AI
IMPACT This research offers a method to enhance the training and inference consistency of flow-based generative models, potentially improving their performance in tasks like text-to-image and text-to-video generation.
RANK_REASON The cluster describes a new research paper detailing a novel method for improving generative models.
Read on Hugging Face Daily Papers →
- Euler--Maruyama discretization
- Flux
- Grpo
- HunyuanVideo
- Langevin correction
- LC-GRPO
- SD3.5-Medium
- Wasserstein error
- ordinary differential equation
- stochastic differential equation
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →
