Gen-Searcher: Search-augmented agent for image generation ( Model and SFT-model on huggingface 8B)

A research team from the Chinese University of Hong Kong (CUHK), UC Berkeley, and UCLA has introduced Gen-Searcher, a novel multimodal AI agent designed to solve a fundamental problem in image generation: outdated or missing knowledge in model parameters. Standard text-to-image models like Stable Diffusion or DALL-E rely solely on their parametric memory, which can be incomplete or stale. Gen-Searcher acts as a front-end agent that first performs multi-hop web searches to gather current textual facts and retrieves relevant reference images, then synthesizes this information into a grounded prompt for an image generator. This search-augmented generation (SAG) approach allows it to handle queries requiring recent events, specific visual details, or niche knowledge that base models lack.

The team developed a dedicated data pipeline to create two training datasets: Gen-Searcher-SFT-10k and Gen-Searcher-RL-6k. They trained the 8B-parameter model first with supervised fine-tuning, followed by agentic reinforcement learning using both text-based and image-based rewards. To evaluate its capability, they introduced KnowGen, a new benchmark specifically for search-dependent image generation. When paired with the Qwen-Image generator, Gen-Searcher improved performance by approximately 16 points on KnowGen and 15 points on the existing WISE benchmark. Crucially, the approach demonstrates transferability, meaning the agent's search-and-grounding logic can enhance other image generators beyond the one it was trained with. The entire project, including the model weights, is fully open-sourced on Hugging Face.