Source-linked AI summary
Gen-Searcher: Reinforcing Agentic Search for Image Generation
Kaituo Feng, Manyuan Zhang, Shuang Chen, Yunlong Lin, Kaixuan Fan, Yilei Jiang, Hongyu Li, Dian Zheng, Chenyang Wang, Xiangyu Yue
TL;DR
Image generators often lack the current, external knowledge needed for complex real-world prompts. Gen-Searcher trains a search-augmented agent with curated data, SFT, and dual-feedback agentic RL, improving performance across KnowGen and WISE while transferring across generators.
Problem
Frozen internal knowledge leaves image generators struggling with knowledge-intensive or up-to-date real-world prompts that require external, sometimes multi-hop, search.
Method
Gen-Searcher uses a dedicated pipeline and two datasets, then trains with SFT followed by GRPO-based agentic RL combining text and image rewards.
Results
Gen-Searcher improves Qwen-Image by around 16 points on KnowGen and around 15 points on WISE, with gains also transferring to other image generators.
Takeaways & Limitations
The work provides an open foundation for search agents in image generation through released data, models, and code.
Takeaways & Limitations
The image generator remains fixed during training, while the search model is optimized to produce grounded prompts and reference images.
Abstract
from arXiv · showhide
Recent image generation models have shown strong capabilities in generating high-fidelity and photorealistic images. However, they are fundamentally constrained by frozen internal knowledge, thus often failing on real-world scenarios that are knowledge-intensive or require up-to-date information. In this paper, we present Gen-Searcher, as the first attempt to train a search-augmented image generation agent, which performs multi-hop reasoning and search to collect the textual knowledge and reference images needed for grounded generation. To achieve this, we construct a tailored data pipeline and curate two high-quality datasets, Gen-Searcher-SFT-10k and Gen-Searcher-RL-6k, containing diverse search-intensive prompts and corresponding ground-truth synthesis images. We further introduce KnowGen, a comprehensive benchmark that explicitly requires search-grounded external knowledge for image generation and evaluates models from multiple dimensions. Based on these resources, we train Gen-Searcher with SFT followed by agentic reinforcement learning with dual reward feedback, which combines text-based and image-based rewards to provide more stable and informative learning signals for GRPO training. Experiments show that Gen-Searcher brings substantial gains, improving Qwen-Image by around 16 points on KnowGen and 15 points on WISE. We hope this work can serve as an open foundation for search agents in image generation, and we fully open-source our data, models, and code.
1 Introduction
Gen-Searcher addresses the difficulty of generating images for knowledge-intensive or up-to-date real-world prompts by combining multi-hop search with image generation. The paper introduces training resources and reports substantial gains on KnowGen and WISE.
- Motivation: Frozen internal knowledge limits image generators on knowledge-intensive prompts involving evolving real-world entities.Such prompts may require multi-hop web search because the necessary information is not always available from a single source.
- Motivation: Static-database retrieval and shallow single-round similarity search remain inadequate for complex, evolving real-world queries.Prior RAG approaches are constrained by database coverage and freshness.
- Approach: Gen-Searcher is presented as the first trained multimodal deep-search agent for image generation.It actively performs multi-hop search and reasoning to gather knowledge for grounded generation.
- Approach: SFT teaches basic tool use, while agentic RL with GRPO optimizes multi-step search trajectories and grounded prompts.The two stages target tool use, browsing, reasoning, and higher-quality search-grounded generation.
- Contributions: The project contributes Gen-Searcher, two training datasets, and the KnowGen benchmark for search-grounded image generation.The project is fully open-sourced as a foundation for future research.
- Results: Around 16 points on KnowGen and around 15 points on WISE are reported as improvements from Gen-Searcher.These results are presented as evidence of effectiveness across knowledge-intensive image-generation evaluations.
2 Related Works
The related work situates Gen-Searcher within progress in image generation and the emergence of agentic reinforcement learning for multi-step reasoning and tool interaction.
- Image Generation: Modern image generation research has progressed from GAN-based methods to diffusion models and increasingly capable systems.Examples include Stable Diffusion, Imagen, FLUX, Qwen-Image, LongCat-Image, Z-Image, and Nano Banana Pro.
- Agentic Reinforcement Learning: Agentic reinforcement learning trains language-model agents for multi-step reasoning and interaction with external tools and environments.Recent work explores long-horizon trajectories and tool-use optimization.
3 Method
The method builds search-grounded image-generation data and evaluates it with KnowGen before training Gen-Searcher through SFT and dual-feedback agentic RL. The design combines external evidence gathering with text- and image-based supervision.
- 3.1 Dataset Construction: The data pipeline constructs search-intensive prompts, agentic trajectories, grounded prompts, reference images, and synthesis ground truth.It addresses the absence of naturally aligned training data for search-grounded image generation.
- 3.1 Dataset Construction: Prompt construction combines Gemini-generated multi-hop prompts across diverse categories with conversions of deep-research questions into image-generation prompts.The complementary strategy primarily expands coverage of general-news scenarios.
- 3.1 Dataset Construction: Agentic trajectories use multi-turn search tools to gather evidence and visual references while providing supervision for supervised fine-tuning.The trajectories support production of a final search-grounded prompt and selected reference images.
- 3.1 Dataset Construction: Approximately 30K raw samples are synthesized, scored for quality, and filtered to obtain approximately 17K high-quality samples.Filtering considers search necessity, correctness, faithfulness, aesthetics, text rendering, safety, token length, and search consistency.
- 3.2 KnowGen Benchmark: KnowGen contains 630 manually verified samples requiring external knowledge and, often, multi-hop evidence aggregation.Its categories span Science & Knowledge and Pop Culture & News.
- 3.2 KnowGen Benchmark: K-Score evaluates faithfulness, visual correctness, text accuracy, and aesthetics using GPT-4.1 judgments and a weighted combination.The weighting assigns 0.4 each to visual correctness and text accuracy, and 0.1 each to faithfulness and aesthetics.
- 3.3 Training Scheme: Gen-Searcher is trained as a multimodal deep-search agent through SFT followed by agentic RL.The downstream image generator remains fixed while the search model produces grounded prompts and reference images.
- 3.3 Training Scheme: Dual reward feedback combines text-based supervision of gathered information with image-based evaluation of final generation performance.The reward is R = (1 −α)Rimage + αRtext, with α = 0.5 and K-Score used as Rimage.
4 Experiments
Gen-Searcher improves search-grounded image generation across KnowGen, WISE, and multiple image-generation backbones. Experiments also show that dual rewards, visual and textual grounding, and learned search behavior contribute to performance, while generator limitations remain.
- Setup: KnowGen evaluates models using Visual correctness, Text accuracy, and an overall K-Score averaged across Science & Knowledge and Pop Culture & News.The benchmark targets real-world, knowledge-intensive prompts requiring external search and multi-step evidence aggregation.
- Main Results on KnowGen: 16.54 points: Gen-Searcher raises Qwen-Image’s KnowGen K-Score from 14.98 to 31.52.It also transfers to Seedream 4.5, improving the score from 31.01 to 47.29, and to Nano Banana Pro, improving it from 50.38 to 53.30.
- Analysis of Different Dimensions: Visual correctness and text accuracy drive most KnowGen gains, while aesthetics can decline slightly when multiple references must be integrated.For Nano Banana Pro, improvement mainly comes from visual correctness because it lacks visual-reference retrieval.
- Performance on WISE: 0.15: Gen-Searcher improves Qwen-Image’s WISE score from 0.62 to 0.77.The largest category gain reported is in Chemistry, from 0.40 to 0.75.
- Ablation Study: 31.52: the full Gen-Searcher outperforms prompt-based workflow at 22.91 and SFT-only training at 28.15 on KnowGen.The ablation attributes additional gains beyond SFT to agentic RL; removing either text or image reward reduces performance.
- Reward Analysis: Both text and image rewards are necessary: setting α to 0 or 1.0 causes clear performance degradation.The dual-feedback design is evaluated by varying the balancing coefficient α during RL training.
5 Conclusion
The paper presents Gen-Searcher as a search agent for knowledge-intensive image generation, supported by dedicated datasets, the KnowGen benchmark, and two-stage training. Experiments report gains across image-generation backbones and benchmarks, with transferability across generators.
- Gen-Searcher is presented as a multimodal deep search agent for knowledge-intensive image generation with agentic reinforcement learning.
- The work introduces Gen-Searcher-SFT-10k, Gen-Searcher-RL-6k, KnowGen, and K-Score for training and evaluating search-grounded image generation.
- Gen-Searcher is trained with supervised fine-tuning followed by agentic reinforcement learning using dual reward feedback.
- The method improves performance across different image-generation backbones on both KnowGen and WISE.
- The trained agent transfers across image generators, including backbones beyond the one used during training.
A KnowGen Benchmark Evaluation Prompt
The KnowGen evaluation prompt defines strict, multi-dimensional scoring for grounded image generation. It separates prompt adherence, ground-truth visual agreement, readable-text accuracy, and aesthetics using discrete scores.
- The evaluator receives a task prompt, generated image, and ground-truth reference image for assessment.
- Each dimension uses only 0, 0.5, or 1, with the top score reserved for fully satisfying the dimension’s criteria.
- K-Score evaluates faithfulness, visual_correctness, text_accuracy, and aesthetics as separate dimensions.
- faithfulness: Faithfulness measures whether the generated image contains the requested subjects, setting, props, actions, relations, counts, and style or format.
- visual_correctness: Visual correctness requires agreement with the reference image’s stable identity and visual features rather than merely matching a generic role archetype.
- text_accuracy: Readable text receives no partial credit when required text is missing, unclear, illegible, or incorrect.
B Text Reward Prompt
The text reward prompt evaluates whether a generated prompt and selected references would support reproducing the ground-truth image. It uses a five-level score focused on textual and visual alignment with the task and reference.
- The evaluator judges a generation prompt and reference-image choices instead of an already generated image.
- Evaluation checks whether the generation prompt captures task and reference requirements and whether selected images provide appropriate grounding.
- The rationale must list hard constraints from textual and visual perspectives before explaining the score.
- The score uses exactly 0, 0.25, 0.5, 0.75, or 1.0, ranging from poor support to fully sufficient alignment.
C System Prompt
The system prompt instructs an agent to search for textual knowledge and visual references, then produce a grounded image-generation prompt and compact reference list. It imposes strict tool-use, identifier, ordering, and deduplication rules.
- The agent searches for missing world knowledge and visual references before producing a generation-ready prompt.
- Search tools: Text search verifies facts such as identities, events, dates, and locations, while image_search grounds visual details such as people, objects, landmarks, and outfits.
- Reference selection: The agent normally selects one image per image-search call and adds more only for distinct, necessary essentials.
- Reference selection: Duplicate references sharing a person, object, event, or venue must be reduced to one clearest and most informative image.
- Output format: The final generation prompt must refer to selected images by ordinal phrases, while the reference list uses IMG_### identifiers.
- Output format: Reference images must be sorted by ascending image identifier so ordinal phrases map unambiguously to list positions.