Read It Back: How Pretrained MLLMs Became Zero-Shot Reward Models for Image Generation

Hugging Face Papers · 2026-07-16

Image: Hugging Face Papers

A paper from ByteDance Seed shows that pretrained multimodal LLMs (MLLMs) work as zero-shot reward models for text-to-image generation.

Instead of training a separate scoring model, the MLLM 'reads back' the generated image and judges how well it matches the prompt.

The approach is appealing because it reuses models that already exist, avoiding costly reward-model training.

For users, better reward signals mean images that follow prompts more faithfully — fewer twisted hands, wrong objects, or ignored details.

The work reflects a broader shift: leaning on general-purpose multimodal models for evaluation rather than narrow specialized ones.