Reinforcing Agentic Search for Image Generation

[🌐 Project Page]    [📖 Paper]

[🤗 GenSearcher-8B-model]    [🤗 SFT-model]    [🤗 GenSearcher-train-data]    [🤗 KnowGen-Bench]

👀 Intro

We introduce Gen-Searcher, as the first attempt to train a multimodal deep research agent for image generation that requires complex real-world knowledge. Gen-Searcher can search the web, browse evidence, reason over multiple sources, and search visual references before generation, enabling more accurate and up-to-date image synthesis in real-world scenarios.

We build two dedicated training datasets Gen-Searcher-SFT-10k, Gen-Searcher-RL-6k and one new benchmark KnowGen for search-grounded image generation.

Gen-Searcher achieves significant improvements, delivering 15+ point gains on the KnowGen and WISE benchmarks. It also demonstrates strong transferability to various image generators.

All code, models, data, and benchmark are fully released.

🔥 News

• [2026/03/30] We release the code, model, data of Gen-Searcher

🔍 KnowGen-Bench

Our KnowGen bench covers around 20 diverse categories in real-world scenarios.

🏆 Performance

Our method delivers consistent gains across backbones, improving Qwen-Image by around 16 points on KnowGen. It also shows strong transferability, generalizing to Seedream 4.5 and Nano Banana Pro with no additional training, yielding about 16-point and 3-point improvements, respectively.

🎥 Demo

Inference Process Example

Application Demo

For more examples, please refer to our website [🌐Project Page]

🚀 Training

Setup

# 1. Clone the repository
git https://github.com/tulerfeng/Gen-Searcher.git
cd Gen-Searcher

# build SFT environment
conda create -n llamafactory python=3.11 
conda activate llamafactory
cd Gen-DeepResearch-SFT/LLaMA-Factory
pip install -e ".[torch,metrics]" --no-build-isolation

# build RL environment
conda create -n genrl python=3.11 
conda activate genrl
cd Gen-DeepResearch-RL

# Install verl
cd rllm/verl
pip install -e .

# Install rllm
cd ../rllm
pip install -e .

# Install diffusers
pip install diffusers
# Finally, install the sglang and vllm

Download the dataset from Hugging Face and unzip all files. Then update the image paths in the JSON file to match your local directory.

SFT Training

Our SFT training simply follow LLaMA-Factory, using our sft_data.json

cd Gen-DeepResearch-SFT/LLaMA-Factory/examples/train_full
llamafactory-cli train gen_qwen3_sft.yaml

A minimum of 8 × 80GB GPUs is required for SFT training based on Qwen/Qwen3-VL-8B-Instruct.

RL Training

First, register the rl_data.json dataset

cd Gen-DeepResearch-RL/rllm/vision_deepresearch_async_workflow/data_prepare
bash register_gen_rl_dataset.sh

Launch the webpage summary model for the browse tool, here we use Qwen/Qwen3-VL-30B-A3B-Instruct

cd Gen-DeepResearch-RL/rllm/vision_deepresearch_async_workflow/run
bash serve_summary_model_vllm.sh

Launch the image generator for rollout generation, here we use Qwen/Qwen-Image-Edit-2509 with FastAPI for serving.

cd qwen_image_api_server/run_server.bash
bash run_server.bash

You may also change to other image generators by simply revising qwen_image_api_server/qwen-image-edit/api.py

Then, set the launched URL serve endpoints for QWEN_EDIT_APP_URL and BROWSE_SUMMARY_BASE_URL in.env.gen_image, and enter your SERPER_KEY_ID, JINA_API_KEYS and GEN_REWARD_API_KEY

It is worth noting that, for open-source release, we replace our original search and image_search tools with the public Serper service. As a result, there may be a performance gap due to differences between the tools. You may also adapt it to your own tools by revising ./rllm/vision_deepresearch_async_workflow/tools

Finally, start the RL training based on the SFT model.

cd Gen-DeepResearch-RL/rllm/vision_deepresearch_async_workflow/run
bash gen_image_deepresearch_8B_fsdp_8gpu.sh

A minimum of 4 × 80GB GPUs is required for RL training.

🔮 Inference

Download the [🤗 KnowGen-Bench], or you may use your own data by setting the same json format. For inference, you also need to properly set the tools and launch the summary model aforementioned.

Serve the trained model

cd Gen-DeepResearch-RL/rllm/vision_deepresearch_async_workflow/run
bash serve_infer_model_vllm.sh

Run the inference script for deep agentic search.

cd Gen-DeepResearch-RL/rllm/eval
bash run_gen_image_eval.sh

Run the generation script for the final image generation based on the search-grounded results

bash run_gen_image_from_results.sh

📐 KnowGen Bench Evaluation

Unzip all files in the KnowGen benchmark.

Set the OpenAI API key and run

cd KnowGen_Eval
bash gpt_eval_knowgen.sh

If you have correctly run the previous agentic inference on KnowGen, the result JSON and generated images should already be properly saved. You only need to ensure that RESULTS_JSON points to the same path as the inference results.

Otherwise, if you use images generated by other methods on KnowGen, you can organize them in the following format for evaluation.

  [
    {
      "id": 3260,
      "success": true,
      "prompt": "xxxxx",
      "meta": { "category": "Biology", "difficulty": "easy" },
      "output_path": "./images/output_3260.png", # Generated image path
      "gt_image": "./gt_image/answer_3260.png" # Ground truth image path
    }
  ]

Acknowledgements

We sincerely appreciate the contributions of the open-source community. The related projects are as follows: verl, LLaMA-Factory, Vision DeepResearch, rllm

Citations

If you find our work helpful for your research, please consider citing our work.