SpectraReward uses MLLM prompt log-likelihood for image generation RL
July 12, 2026
SpectraReward enables training-free reinforcement learning for text-to-image models by using a pretrained MLLM's image-conditioned prompt log-likelihood as a reward signal. This method reuses existing image-text alignment capabilities without requiring preference labels or fine-tuning a separate reward model.
HOW THIS AFFECTS YOU
●
builderYou can implement more efficient RLHF pipelines for generative models without training dedicated reward models.
●
researcherYou can bypass the need for preference dataset collection by utilizing MLLM log-likelihoods for reward modeling.