Runhui Huang, Qihui Zhang, Zhe Liu, Yu Gao, Jie Wu, Hengshuang Zhao
4 min
Abstract
In this paper, we propose SpectraReward, a training-free reward function that turns pretrained MLLMs into off-the-shelf reward models for image-generation reinforcement learning. Instead of asking the MLLM to judge a generated image or answer decomposed verification questions, SpectraReward measures how well the original prompt can be recovered from the generated image through a single image-conditioned, teacher-forced forward pass. We use the average image-conditioned prompt log-likelihood as the reward, directly reusing the MLLM's pretrained image-text alignment ability without preference labels, reward-model fine-tuning. We further introduce Self-SpectraReward, a special case for unified multimodal models where the policy's own understanding branch serves as the reward model for its generation branch, forming a closed-loop self-improving framework without external reward models or external knowledge. Extensive experiments validate SpectraReward through a broad image-generation RL study covering two diffusion models, three RL algorithms, nine reward MLLM backbones from four MLLM families spanning 4B to 235B parameters, and five out-of-distribution text-to-image benchmarks. Results show that both SpectraReward and Self-SpectraReward significantly and consistently improve generation performance and outperform prior MLLM-derived reward training methods. Further analysis reveals that larger reward MLLMs are not always better, while Self-SpectraReward can match or surpass much larger external reward models, suggesting that reward-policy alignment is a key factor for effective image-generation RL. Project Page: https://huangrh99.github.io/SpectraReward/
Sam: So it's like a student who also wrote the marking scheme. They know exactly what a correct answer looks like, because they built the criteria themselves.
Alex: And that shared background turns out to matter. Because both sides were trained together on the same data, they speak the same internal language. The paper suggests this alignment — the evaluator and generator being in sync — is a key reason the self-grading approach often outperforms using an external judge.
Sam: Does that actually show up in the results, or is it more of a theoretical argument?
Alex: The study reports it as a clear improvement across multiple standard benchmarks — consistently outperforming methods that rely on human-labelled training data. The paper frames this as evidence that internal alignment matters, not just model size or raw capability.
Sam: Though I'd imagine there are limits. If the prompt is abstract — something like "a melancholy afternoon" — there's no single visual fact to check for.
Alex: That's a real constraint the paper acknowledges. The method works best with concrete, literal descriptions. If a prompt implies something rather than stating it — like "hot coffee" suggesting steam should be present — the system may not catch that gap, because it's only checking for what was explicitly written.
Sam: So it's precise and scalable for literal instructions, but it still needs human oversight for anything more interpretive.
Alex: That's a fair summary. It's a meaningful step toward models that can evaluate their own output without expensive human feedback — but it's a tool with a defined scope, not a complete solution. Thanks for listening to ResearchPod.