Beyond VLM-Based Rewards: Diffusion-Native Latent Reward Modeling
Researchers have developed DiNa-LRM, a novel diffusion-native latent reward model designed to improve preference learning for diffusion and flow-matching models. This new approach formulates preference learning directly on noisy diffusion states, overcoming the domain mismatch issues associated with using Vision-Language Models (VLMs) for reward provision. DiNa-LRM offers competitive performance to state-of-the-art VLMs but at a significantly reduced computational cost, leading to faster and more efficient model alignment. AI
IMPACT Introduces a more computationally efficient method for aligning diffusion models, potentially accelerating their development and application.