---
license: apache-2.0
language:
- en
library_name: transformers
pipeline_tag: image-text-to-text
base_model: Qwen/Qwen3-VL-8B-Instruct
tags:
- agent
- image-generation
- tool-use
- visual-reasoning
- self-distillation
- grpo
- reinforcement-learning
- multimodal
- qwen3-vl
datasets:
- MeiGen-AI/GenEvolve-Data-Bench
---

<div align="center">

<img src="assets/logo_genevolve.png" alt="GenEvolve" width="160">

<h1>GenEvolve</h1>

<p><strong><em>Self-Evolving Image Generation Agents via Tool-Orchestrated Visual Experience Distillation</em></strong></p>

<p>
  <a href="https://arxiv.org/abs/2605.21605">
    <img alt="Paper" src="https://img.shields.io/badge/📄_Paper-arXiv:2605.21605-b31b1b"></a>
  <a href="https://ephemeral182.github.io/GenEvolve/">
    <img alt="Project Page" src="https://img.shields.io/badge/🌐_Project-Page-1f6feb"></a>
  <a href="https://github.com/MeiGen-AI/GenEvolve">
    <img alt="Code" src="https://img.shields.io/badge/💾_GitHub-Code-181717"></a>
  <a href="https://huggingface.co/datasets/MeiGen-AI/GenEvolve-Data-Bench">
    <img alt="Dataset" src="https://img.shields.io/badge/🤗_Dataset-GenEvolve--Data-FFD21E"></a>
</p>

</div>

This repository hosts the **GenEvolve agent policy** — a Qwen3-VL-8B-Instruct backbone fine-tuned and self-evolved into a tool-orchestrated image-generation agent. Given a user request, the agent issues web/image searches, retrieves visual references, activates internal generation knowledge, and emits an executable **prompt-reference program** `z = (gen_prompt, reference_images)` that drives any reference-conditioned downstream generator (Qwen-Image-Edit, Nano Banana Pro, ...).

<div align="center">
<img src="assets/teaser.jpg" alt="GenEvolve teaser" width="100%">

<p><em>The same trained agent policy paired with two reference-conditioned generators ⟶<br>
<strong>Qwen-Image-Edit (open)</strong> &nbsp;·&nbsp; <strong>Nano Banana Pro (strong)</strong></em></p>
</div>

---

## ✨ 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)

| Method | Generator | KScore | Knowledge-Anch. | Quality-Anch. |
|---|---|---:|---:|---:|
| Qwen-Image (raw) | Qwen-Image | 0.2987 | 0.2384 | 0.3768 |
| Nano Banana Pro (raw) | Nano Banana Pro | 0.5298 | 0.5160 | 0.5477 |
| Gen-Searcher 8B | Qwen-Image-Edit-2511 | 0.3493 | 0.3293 | 0.3745 |
| Gen-Searcher 8B | Nano Banana Pro | 0.5481 | 0.5472 | 0.5492 |
| **GenEvolve (Ours)** | Qwen-Image-Edit-2511 | **0.3663** | **0.3410** | **0.3990** |
| **GenEvolve (Ours)** | Nano Banana Pro | **0.5739** | **0.5669** | **0.5830** |

### WISE Benchmark (WiScore, six knowledge categories)

| Model | Cultural | Time | Space | Biology | Physics | Chemistry | **Overall** |
|---|---:|---:|---:|---:|---:|---:|---:|
| GPT-4o | 0.81 | 0.71 | **0.89** | **0.83** | 0.79 | 0.74 | 0.80 |
| Gen-Searcher-8B + Qwen-Image | 0.80 | 0.71 | 0.82 | 0.76 | 0.74 | 0.75 | 0.77 |
| Mind-Brush | 0.83 | 0.69 | 0.84 | 0.71 | **0.85** | 0.68 | 0.78 |
| **GenEvolve + Qwen-Image-Edit** | **0.84** | 0.74 | 0.87 | **0.83** | 0.81 | **0.83** | **0.82** |

---

## 🧠 Method Overview

<p align="center"><img src="assets/overview.png" alt="GenEvolve method overview" width="92%"></p>

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

<p align="center"><img src="assets/visual_comparison.png" alt="Qualitative comparison" width="100%"></p>

<p align="center"><sub>Qualitative comparison on representative cases. <span style="color:#D97706">Orange</span> marks external/uncommon knowledge requirements; <span style="color:#2563EB">blue</span> marks internal generation-knowledge requirements.</sub></p>

### 🎨 Gallery — paired with Nano Banana Pro

<p align="center"><img src="assets/gallery_nano.jpg" alt="GenEvolve + Nano Banana Pro gallery" width="100%"></p>

<p align="center"><sub>The same agent policy with Nano Banana Pro as the downstream renderer. Examples cover spatial layout, text rendering, quantity counting, attribute binding, anatomy/pose, creative transfer, material physics, and aesthetic drawing.</sub></p>

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

<p align="center"><img src="assets/gallery_qwen.jpg" alt="GenEvolve + Qwen-Image-Edit gallery" width="100%"></p>

<p align="center"><sub>Same trained policy paired with the open-source Qwen-Image-Edit-2511 renderer; consistent quality across both generators reflects generator-transferable orchestration.</sub></p>

---

## 🚀 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](https://github.com/MeiGen-AI/GenEvolve); the snippet below mirrors its installation and usage.

### 1. Install the main GenEvolve runtime

```bash
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
# 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
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
{
  "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

| Artifact | Link |
|---|---|
| Project page | https://ephemeral182.github.io/GenEvolve/ |
| Paper | Coming soon |
| Code | https://github.com/MeiGen-AI/GenEvolve |
| Training data + benchmark | [MeiGen-AI/GenEvolve-Data-Bench](https://huggingface.co/datasets/MeiGen-AI/GenEvolve-Data-Bench) |
| Base model | [Qwen/Qwen3-VL-8B-Instruct](https://huggingface.co/Qwen/Qwen3-VL-8B-Instruct) |

---

## ⚖️ 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
@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}, 
}
```