GenEvolve - A self-evolving image generation agent launched by Meituan and others

What is GenEvolve?

GenEvolve is a self-evolving image generation agent jointly developed by the Hong Kong University of Science and Technology (Guangzhou), Meituan, and the National University of Singapore. GenEvolve formalizes open-ended image generation as a visual trajectory orchestrated by tools. The agent searches for textual evidence, retrieves visual references, and invokes generative knowledge to ultimately synthesize a prompt-reference program, which is then rendered by an arbitrary reference condition generator.

GenEvolve's main functions

• Text search (search)Collect external factual evidence to supplement global knowledge of real buildings, public figures, product structures, etc.
• Image search (image_search)Search for visual reference images to obtain visual references such as landmarks, people, and materials.
• Generate knowledge query (query_knowledge)Activate internal generation skills to handle complex requirements such as text rendering, spatial layout, and material consistency.
• Prompt-Reference Program synthesisIt integrates tool results into an executable generator that supports rendering of generators with arbitrary reference conditions.

GenEvolve's technical principles

• The training is divided into two phases.:
  • SFT Cold StartSupervised fine-tuning of Qwen3-VL-8B-Instruct on a high-quality Teacher trajectory, learning basic tool calls and program writing.
  • Rollout (self-evolution)Multiple trajectories are sampled for the same request, rendered into images, and scored by both a visual scorer and a text scorer. GRPO is used to optimize trajectory-level rewards.
• Visual Experience DistillationBy comparing the best and worst trajectories under the same request, the differences are summarized into a structured Decision Guide. The preferences of the teacher branch are distilled into the student model through token-level reverse KL, so that the model learns the decision-making habits of "how to search, select references and organize constraints when similar requests are made".

Follow us on WeChat and reply with "open source",join inAI open source project discussion group

How to use GenEvolve

• Cloning repositoryDownload the GenEvolve code from GitHub to your local environment.
• Install dependencies:run pip install -r requirements.txt Install the required Python packages.
• Download ModelDownload from HuggingFace MeiGen-AI/GenEvolve The Qwen3-VL-8B Agent policy weights.
• Configure search toolSet the search engine API key and enable it. search and image_search Tool call.
• Loading and generating knowledgePlace the eight skills Markdown files in the specified directory for later use. query_knowledge Tool reading.
• Select generatorSelect the backend and configure the key in Qwen-Image-Edit or Nano Banana Pro.
• Initialize AgentInstantiation GenEvolveAgentPass in the model path, generator name, and tool list.
• Input Request: call agent.run() The method takes a natural language image as input and generates a description.
• Execution trajectoryThe agent automatically searches for textual evidence, retrieves visual references, invokes generated knowledge, and synthesizes a prompt-reference program.
• Rendering imagesThe program output by the Agent is passed to the configured generator backend to obtain the final generated image.

GenEvolve's core advantages

• Generator portabilityThe same agent strategy can be used with the open-source Qwen-Image-Edit or the strong generator Nano Banana Pro, and the strategy is not bound to a specific generator.
• Self-evolution closed loopGRPO + visual experience self-distillation forms a closed loop of "stronger strategy → better trajectory → richer experience → future improvement".
• Fine-grained credit allocationCompared to the scalar rewards of pure RL, SDL provides token-level guidance, clarifying whether a good trajectory is a good tool plan, a reference selection, or a constraint writing method.
• Cross-benchmark generalizationWithout in-domain fine-tuning on the WISE knowledge-intensive benchmark, the 8B open-source strategy surpasses GPT-4o.

GenEvolve's project address

• Project official website: https://ephemeral182.github.io/GenEvolve/
• GitHub repositoryhttps://github.com/MeiGen-AI/GenEvolve
• HuggingFace model libraryhttps://huggingface.co/MeiGen-AI/GenEvolve

GenEvolve's Competitive Comparison

Application scenarios of GenEvolve

• Knowledge-intensive image generationFor scenarios requiring knowledge of the external world, such as real landmarks (Eiffel Tower, Forbidden City), portraits of public figures, specific product structures, and historical event reconstructions, the Agent supplements the facts using search tools to avoid the generator's "illusion" of incorrect details.
• Quality-constrained image generationCommercial designs, advertising posters, and educational illustrations that have strict requirements for text rendering, accurate counting, spatial layout, attribute binding, anatomical accuracy, material realism, and aesthetic style.
• Reference Consistency Image GenerationThe requirements are to produce a series of IP derivative works that maintain consistency with the reference images (character design), use special materials (metal/silk texture) or have a unified style (specific painter's style).
• Clarification and proceduralization of ambiguous requirementsTransform vague user descriptions into executable programs, clearly defining what to search, what to reference, and what to constrain, thus lowering the barrier to prompt engineering.