GenEvolve

✶Self-Evolving Image Generation Agents via
Tool-Orchestrated Visual Experience Distillation✶

Sixiang Chen^1,2, Zhaohu Xing¹, Tian Ye¹, Xinyu Geng³, Yunlong Lin, Jianyu Lai^1,2, Xuanhua He³, Fuxiang Zhai¹, Jialin Gao^4,‡, Lei Zhu^1,3,†

¹The Hong Kong University of Science and Technology (Guangzhou) ²Meituan ³The Hong Kong University of Science and Technology ⁴National University of Singapore Project Leader: Junfeng Luo (Meituan)

Paper Code 🤗 Model 🤗 Dataset Gallery

🎨 Representative images generated by GenEvolve agents with Nano Banana Pro & Qwen-Image-Edit — spanning architecture, nature, creative transfer, street scenes, scientific illustration, and more.

Abstract

Open-ended image generation is no longer a simple prompt-to-image problem. High-quality generation often requires an agent to combine a model's internal generative ability with external resources. As requests become more diverse and demanding, we aim to develop a general image-generation agent that can self-evolve through trajectories and use tools more effectively across varied generation challenges.

To this end, we propose GenEvolve, a self-evolving framework based on Tool-Orchestrated Visual Experience Distillation. In GenEvolve, each generation attempt is modeled as a tool-orchestrated trajectory, where the agent gathers evidence, selects references, invokes generation skills, and composes them into a prompt-reference program. Unlike existing methods that rely on image-level scalar rewards, GenEvolve compares multiple trajectories and abstracts best-worst differences into structured visual experience, provided only to a privileged teacher branch. Inspired by on-policy self-distillation, Visual Experience Distillation provides dense token-level supervision for better tool orchestration.

We further construct GenEvolve-Data and GenEvolve-Bench. Experiments show substantial gains over strong baselines, achieving state-of-the-art performance among current image-generation frameworks.

Contributions

Key Highlights

Tool-Orchestrated Trajectories

Models generation as multi-turn visual trajectories with external search, visual references, and callable generation knowledge.

Visual Experience Distillation

Converts best-worst trajectory differences into structured experience, enabling dense token-level supervision via privileged teacher distillation.

Self-Evolving Loop

GRPO + VED form a closed loop: stronger policies → better trajectories → richer experience → future improvements.

GenEvolve-Data & Bench

Comprehensive trajectory dataset and diagnostic benchmark spanning Knowledge-Anchored and Quality-Anchored challenges.

Methodology

Framework Overview

Overview of GenEvolve. The student agent orchestrates search, references, and generation knowledge to produce a prompt-reference program. Multiple trajectories are judged; best-worst differences form visual experience injected into the teacher. GRPO + Visual Experience Self-Distillation closes the self-evolving loop.

Tool-Orchestrated Trajectories

The agent samples trajectories with three tool families: search(q) for textual evidence, image_search(q) for visual references, and query_knowledge(skill) for callable generation skills (text rendering, layout, anatomy, aesthetics, material, counting, etc.)

Prompt-Reference Program Synthesis

The trajectory outputs z=(g,R) — a targeted instruction with ordinal references, binding constraints from the user request, retrieved facts, selected references, and activated generation knowledge.

Visual Experience Extraction

Best-worst trajectory pairs are summarized into five structured experience slots: search strategy, knowledge activation, reference selection, prompt construction, and failure avoidance.

Visual Experience Self-Distillation

The teacher receives experience-patched context; student sees only normal context. Importance-weighted reverse-KL distillation provides dense token-level guidance without changing inference.

Dataset

GenEvolve-Data & GenEvolve-Bench

A complete trajectory-level dataset for training and evaluating image-generation agents — from diverse prompts through tool-orchestrated teacher trajectories to filtered GT images.

Data construction pipeline. Diverse prompts → tool-orchestrated teacher trajectories → VLM audit → GT image rendering & filtering → splits for SFT cold-start, self-evolution, and held-out evaluation.

📋

Prompt Pool

20K structured recipes spanning Knowledge-Anchored (entities, events, places) and Quality-Anchored (text, layout, counting, anatomy, material, aesthetics) generation challenges.

🔄

Teacher Trajectories

Each prompt is solved through a real multi-turn tool loop by strong teacher models (Seed2.0, Gemini 3 Pro), recording search queries, selected references, activated skills, and final programs.

🔍

Filtering & Audit

Programmatic checks + VLM judge remove incomplete tool loops, invalid reference selections, and underspecified programs. Only high-quality trajectories survive.

🎯

GT Images & Splits

Teacher programs are rendered into GT images by Nano Banana Pro, filtered for quality, then split into SFT (trajectories), self-evolution (image feedback), and GenEvolve-Bench (held-out eval).

📦 Data Examples

Each entry contains a user request, full tool trajectory, selected references, and GT image.

Agent Workflow

Trajectory Visualization

See how GenEvolve orchestrates tools step-by-step: from user request through search, reference selection, skill activation, to the final generated image.

Experiments

Results

Main Comparison on GenEvolve-Bench

Method	Generator	Judge Dimensions				Benchmark Overall
Method	Generator	Faith.	Vis.	Text	Aesth.	KScore	Know.	Qual.
Direct Generator Baselines
Lumina-Image 2.0	Lumina-Image 2.0	0.1044	0.0000	0.3308	0.2694	0.1697	0.1528	0.1915
BAGEL	BAGEL	0.1212	0.0059	0.3721	0.4082	0.2041	0.1684	0.2504
SD-3.5-Large	SD-3.5	0.1456	0.0135	0.3872	0.4865	0.2235	0.1943	0.2612
FLUX.1-dev	FLUX.1	0.1574	0.0059	0.4150	0.5556	0.2396	0.2097	0.2784
FLUX.2 Klein 4B	FLUX.2	0.2525	0.0059	0.3847	0.5648	0.2380	0.2004	0.2865
Z-Image-Turbo	Z-Image	0.2837	0.0396	0.4369	0.6187	0.2808	0.2340	0.3413
FLUX.2 Klein 9B	FLUX.2	0.3662	0.0210	0.4192	0.6599	0.2787	0.2327	0.3382
Z-Image	Z-Image	0.3333	0.0278	0.4352	0.5429	0.2728	0.2203	0.3407
Qwen-Image	Qwen-Image	0.3729	0.0623	0.4226	0.6751	0.2987	0.2384	0.3768
Nano Banana Pro	Nano Banana Pro	0.7761	0.2837	0.6178	0.9158	0.5298	0.5160	0.5477
Agentic Image-Generation Workflows
Gen-Searcher 8B	Qwen-Image-Edit-2511	0.5284	0.1050	0.4768	0.6377	0.3493	0.3293	0.3745
Gen-Searcher 8B	Nano Banana Pro	0.7465	0.3378	0.6198	0.9036	0.5481	0.5472	0.5492
GenEvolve (Ours)	Qwen-Image-Edit-2511	0.5303	0.1338	0.4907	0.6347	0.3663	0.3410	0.3990
GenEvolve (Ours)	Nano Banana Pro	0.7970	0.3832	0.6218	0.9222	0.5739	0.5669	0.5830

External Evaluation on WISE Benchmark

Model	Cultural	Time	Space	Biology	Physics	Chemistry	Overall
Direct Generator Baselines
Emu3	0.34	0.45	0.48	0.41	0.45	0.27	0.39
FLUX.1-schnell	0.39	0.44	0.50	0.31	0.44	0.26	0.40
SD-3-Medium	0.42	0.44	0.48	0.39	0.47	0.29	0.42
SD-3.5-Medium	0.43	0.50	0.52	0.41	0.53	0.33	0.45
SD-3.5-Large	0.44	0.50	0.58	0.44	0.52	0.31	0.46
FLUX.1-dev	0.48	0.58	0.62	0.42	0.51	0.35	0.50
Hunyuan-Image 3.0	0.58	0.57	0.70	0.56	0.63	0.31	0.57
UniWorld-V2	0.60	0.61	0.70	0.53	0.64	0.32	0.58
Qwen-Image	0.62	0.63	0.77	0.57	0.75	0.40	0.62
NextFlow-RL	0.63	0.63	0.77	0.58	0.67	0.39	0.62
LongCat-Image	0.66	0.61	0.72	0.66	0.72	0.49	0.65
DeepGen1.0	0.72	0.81	0.70	0.67	0.82	0.66	0.73
GPT-4o	0.81	0.71	0.89	0.83	0.79	0.74	0.80
Agentic Image-Generation Workflows
GenAgent	0.78	0.67	0.78	0.72	0.77	0.55	0.72
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 (Ours)	0.84	0.74	0.87	0.83	0.81	0.83	0.82

Visual comparison on GenEvolve-Bench. Orange = external knowledge requirements. Blue = internal generation-knowledge requirements.

Showcase

Generation Gallery

Click any card to view references, generated program, and full details.

Citation

BibTeX

@article{chen2026genevolve,
  title   = {GenEvolve: Self-Evolving Image Generation Agents via 
             Tool-Orchestrated Visual Experience Distillation},
  author  = {Chen, Sixiang and Xing, Zhaohu and Ye, Tian and Geng, Xinyu 
             and Lin, Yunlong and Lai, Jianyu and He, Xuanhua and Zhai, Fuxiang 
             and Gao, Jialin and Zhu, Lei},
  journal = {arXiv preprint arXiv:XXXX.XXXXX},
  year    = {2026}
}