✨ Highlights

• Tool-orchestrated trajectories. The agent calls search, image_search, and query_knowledge (8 callable generation skills) before producing a final program z = (gen_prompt, reference_images).
• Self-evolution with Visual Experience Distillation. Best-vs-worst trajectory pairs are distilled token-level into the deployed student. No runtime memory at inference.
• Generator-transferable. The same trained policy works with both an open-source generator (Qwen-Image-Edit-2511) and a strong proprietary generator (Nano Banana Pro).

📊 Headline Results

GenEvolve-Bench (KScore, held-out split)

WISE Benchmark (WiScore, six knowledge categories)

---

🧠 Method Overview

For a user request, the agent samples a multi-turn trajectory of tool calls before emitting the final prompt-reference program. The downstream generator then renders the image.

---

🖼️ Visual Demos

🎨 Gallery — paired with Nano Banana Pro

🎨 Gallery — paired with Qwen-Image-Edit (open)

---

🚀 Quick Start

The deployed checkpoint is the student policy — it consumes a user prompt and returns a JSON gen_prompt + reference_images program through a <think>/<tool_call>/<answer> loop. The end-to-end runtime (vLLM serving + agent loop + tools + Qwen/Nano renderers) lives in the GitHub repo; the snippet below mirrors its installation and usage.

1. Install the main GenEvolve runtime

bash
Copy code
git clone https://github.com/MeiGen-AI/GenEvolve.git
cd GenEvolve

conda create -n genevolve python=3.11 -y && conda activate genevolve
pip install -U pip setuptools wheel packaging psutil ninja
pip install torch==2.8.0 torchvision==0.23.0 --index-url https://download.pytorch.org/whl/cu128
pip install --no-build-isolation -r requirements.txt
pip install -e .

Qwen-Image-Edit rendering runs as a separate FastAPI service (kept out of the vLLM environment to avoid CUDA/diffusers conflicts). Set up that service from the GitHub README when you want to use --backend qwen-image-edit-service.

2. Serve the agent policy

bash
Copy code
# Single GPU / single replica.
MODEL_PATH=MeiGen-AI/GenEvolve PORT=8000 TP=1 DP=1 bash scripts/serve_vllm.sh

# Higher throughput on one 8-GPU node (8 replicas, 1 GPU each).
MODEL_PATH=MeiGen-AI/GenEvolve PORT=8000 TP=1 DP=8 bash scripts/serve_vllm.sh

TP shards one model replica across multiple GPUs; DP launches multiple replicas; total GPU usage is TP × DP.

3. End-to-end example

bash
Copy code
export SERPER_API_KEY=<your_key>      # required for search / image_search
export GOOGLE_API_KEY=<your_key>      # or GEMINI_API_KEY; only for --backend nano-banana-pro

# Nano Banana Pro renderer
python examples/quickstart.py \
    --backend nano-banana-pro \
    --base-url http://localhost:8000/v1 \
    --model GenEvolve \
    --prompt "A 1990s travel-magazine cover of two backpackers in front of the Eiffel Tower at golden hour, the title \"PARIS\" in bold serif." \
    --output paris.png

# Qwen-Image-Edit renderer (point at your Qwen-Image-Edit FastAPI service)
python examples/quickstart.py \
    --backend qwen-image-edit-service \
    --service-url http://your-qwen-service:8001 \
    --base-url http://localhost:8000/v1 \
    --model GenEvolve \
    --output paris_qwen.png

The agent's final <answer> is a JSON object:

json
Copy code
{
  "gen_prompt": "...natural-language prompt that refers to images by 'the first reference image', ...",
  "reference_images": [
    {"img_id": "IMG_001", "note": "what to copy from this image"}
  ]
}

gen_prompt MUST refer to selected images using ordinal phrases ("the first reference image") — never raw IMG_### ids or URLs. Pass (gen_prompt, [r["local_path"] for r in reference_images]) to your favourite reference-conditioned generator (Qwen-Image-Edit, Nano Banana Pro, ...) to obtain the final image.

---

🗂️ Related Artifacts

---

⚖️ Intended Use, Limits, Bias

• Intended use. Research on tool-using image-generation agents, agentic prompt-program synthesis, and self-distillation from generated outcomes.
• Search dependency. The agent issues live web/image queries through user-provided tool wrappers. Quality of grounded facts depends on the search backend you plug in.
• Bias. Tool outputs and reference images come from public web search, which carries demographic, cultural, and geographic biases that may be reflected in agent outputs.

---

📑 Citation

bibtex
Copy code
@misc{chen2026genevolveselfevolvingimagegeneration,
      title={GenEvolve: Self-Evolving Image Generation Agents via Tool-Orchestrated Visual Experience Distillation}, 
      author={Sixiang Chen and Zhaohu Xing and Tian Ye and Xinyu Geng and Yunlong Lin and Jianyu Lai and Xuanhua He and Fuxiang Zhai and Jialin Gao and Lei Zhu},
      year={2026},
      eprint={2605.21605},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2605.21605}, 
}