meigen-ai/genevolve

GenEvolve is a research project from teams at Hong Kong University of Science and Technology and Meituan, published as an academic paper. The goal is to build an AI agent that can generate images in response to open-ended prompts, not by generating pixels directly, but by first gathering information and then producing a structured program that tells a separate image generation model what to create and what reference images to use. The agent works by calling a set of tools: a web search tool to find textual information about the requested subject, an image search tool to find visual references, and a knowledge query tool to access generation skills the agent has learned. After gathering that material, the agent produces a combined output of a text prompt and a set of reference images, which gets passed to a backend image generator. The key claim of the research is that this output is generator-transferable: the same agent output can be sent to the open Qwen-Image-Edit model or to a stronger proprietary image generator, without the agent needing to be retrained for each one. The released agent policy is built on Qwen3-VL-8B, a vision-language model, and is served via vLLM for efficient inference. The repository includes the trained agent weights on HuggingFace, the three tools, eight skill description files used during training, wrappers for both supported image generation backends, and around 9,000 training trajectories plus a 594-prompt evaluation benchmark. Setting up the project requires a CUDA-capable GPU, Python 3.11, and a specific set of library versions including PyTorch 2.8, vLLM 0.11, and Flash Attention. The image generation service runs as a separate process from the agent, and external API keys for Serper (web search) and optionally Google's Generative Language API are needed to use the full tool set. This is a research codebase accompanying an arXiv paper, not a consumer product. The README is detailed and assumes familiarity with GPU model serving and machine learning infrastructure.