yejy53/RealGen

What it solves

RealGen addresses the lack of photorealism in AI-generated images, specifically targeting the removal of artifacts and the enhancement of fine details to make generated images indistinguishable from real photographs.

How it works

RealGen uses a post-training approach called Detector Reward-guided GRPO. It utilizes AIGC (AI-Generated Content) detectors as reward signals; the generator is trained to "escape" detection by the AI detector, which forces the model to reduce artificial patterns and improve physical fidelity.

RealGen-V2 further evolves this by introducing a dynamic GAN-inspired online reinforcement learning paradigm.

Who it’s for

This project is for researchers and developers working on text-to-image generation who want to achieve higher levels of photorealism and reduce AI-generated artifacts.

Highlights

• Detector-Guided Rewards: Uses AI detectors to guide the model toward photorealism via reinforcement learning.
• RealBench: A new benchmark for evaluating photorealism using automated Detector-Scoring and Arena-Scoring, removing the need for human evaluators.
• High Performance: Outperforms general image models and specialized realistic models in aesthetics and detail.
• Framework Compatibility: Compatible with various diffusion-model-based GRPO paradigms, such as Dance GRPO and Flow GRPO.