Research map / Image generation / Text-to-image diffusion
Gen-Searcher: Reinforcing Agentic Search for Image Generation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Trains Gen-Searcher, a Qwen3-VL-8B search agent that performs multi-hop text and image search to write grounded prompts and select reference images for image generators, using SFT then GRPO with a dual text+image reward, and introduces the KnowGen benchmark with K-Score.
Problem
Image generators rely on frozen parametric knowledge and fail on knowledge-intensive or up-to-date prompts that require web search and visual references.
Approach
A search agent trained with agentic RL to iteratively search, browse, and retrieve reference images, producing a search-grounded prompt for a frozen downstream image generator; RL uses a dual reward combining a GPT-4.1 text-sufficiency score with an image-based K-Score.
Key results
- Gen-Searcher-8B + Qwen-Image raises KnowGen K-Score from 14.98 to 31.52 (+16.54).
- On WISE, Gen-Searcher-8B + Qwen-Image improves overall from 0.62 to 0.77.
- Gen-Searcher transfers to Seedream 4.5 (31.01 to 47.29) and Nano Banana Pro (50.38 to 53.30) without retraining.
Surprising findings
- Pure image-based reward is noisy for RL because downstream generator stochasticity dominates; adding a text reward is needed.
- Nano Banana Pro gains come mainly from visual correctness, with text accuracy nearly unchanged.
Limitations
- Relies on proprietary models (Gemini 3 Pro, Nano Banana Pro, Seed1.8) for data generation and GPT-4.1 for judging.
- Downstream open-source generators still limit final quality (e.g., multi-subject consistency and text rendering).
- Aesthetics sometimes decreases slightly after search-grounded prompting.
Experimental setup
Datasets: Gen-Searcher-SFT-10k Gen-Searcher-RL-6k deep research QA datasets
Benchmarks: KnowGen WISE
Metrics: K-Score WISE score
Baselines: Qwen-Image Nano Banana Pro GPT-Image-1.5 Seedream 4.5 FLUX.2-klein-9B HunyuanImage-3.0 LongCat-Image Qwen-Image + workflow
Builds on: GRPO WISE Qwen3-VL Qwen-Image
Closest papers by meaning
- Generation Navigator: A State-Aware Agentic Framework for Image Generation (NeurIPS 2026 · similarity 0.90)
- SenseSearch: Empowering Vision-Language Models with High-Resolution Agentic Search-Reasoning via Reinforcement Learning (CVPR 2026 · similarity 0.88)
- Unify-Agent: A Unified Multimodal Agent for World-Grounded Image Synthesis (NeurIPS 2026 · similarity 0.87)
- T2I-R1: Reinforcing Image Generation with Collaborative Semantic-level and Token-level CoT (NeurIPS 2025 · similarity 0.87)
- OpenSearch-VL: An Open Recipe for Frontier Multimodal Search Agents (NeurIPS 2026 · similarity 0.86)
- GoT-R1: Unleashing Reasoning Capability of Autoregressive Visual Generation with Reinforcement Learning (ICLR 2026 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.