Brief

Researchers have introduced AlphaGRPO, a new framework designed to improve multimodal generation in Unified Multimodal Models (UMMs). This approach uses Group Relative Policy Optimization (GRPO) to enable models to perform advanced reasoning tasks like inferring user intent for text-to-image generation and self-correcting outputs. To provide better supervision, AlphaGRPO incorporates a Decompositional Verifiable Reward (DVReward) system, which breaks down user requests into verifiable questions evaluated by a general multimodal large language model (MLLM). Experiments show AlphaGRPO significantly enhances performance on various multimodal generation and editing benchmarks. AI

Researchers have developed a new framework called Latent-to-Pixel (L2P) that efficiently transfers knowledge from pre-trained Latent Diffusion Models (LDMs) to create powerful pixel-space models. This method avoids the need for extensive computational resources and real-world data by freezing most of the source LDM and training only shallow layers for the latent-to-pixel transformation. L2P utilizes synthetic images generated by LDMs as its training corpus, enabling rapid convergence with minimal hardware. The approach also eliminates the VAE bottleneck, allowing for native generation of ultra-high resolution images. AI

Researchers have developed a new method called Super-Linear Advantage Shaping (SLAS) to improve text-to-image models trained with reinforcement learning. This technique addresses reward hacking by reshaping the policy space using an information geometry perspective, amplifying informative updates while suppressing noisy ones. SLAS demonstrates superior performance over existing methods like DanceGRPO, leading to faster training, better out-of-domain generation, and increased robustness to model scaling. AI