AIAny
Icon for item

ToolArtist: Tool-Using Unified Multimodal Models for Agentic Image Generation

Orchestrates reasoning, external tool use, and native image generation under one unified multimodal agent policy via post-training. Introduces RAD-GRPO for agentic reinforcement fine-tuning and releases training data plus the full post-training infrastructure.

Introduction

Most text-to-image systems separate reasoning, tool use (search, retrieval), and image synthesis into fixed pipelines, which limits open‑world generation that needs multi-step reasoning and external knowledge. The core insight here is that placing the entire open‑world image generation loop under a single unified multimodal agent policy — through post-training a Unified Multimodal Model (UMM) and agentic RL — yields more coordinated, verifiable, and higher-quality results than piecemeal or fixed workflows.

Key Findings
  • Unified agentic control: The work shows how supervised fine-tuning with a teacher agent (that uses search tools and an image-generation tool) can produce UMM trajectories where the image tool is concealed but generated outputs are kept, enabling a single policy to learn when and how to reason, call tools, and produce images.
  • Agentic RL with RAD-GRPO: They introduce Reason-Act-Draw GRPO (RAD-GRPO), an RL scheme combining complementary intent and quality rewards to jointly optimize reasoning/action selection and image quality under one model.
  • Empirical outcome: Experiments indicate that a fully agentic pipeline (reasoning, tool invocation, and native generation coordinated by one policy) consistently outperforms approaches that use fixed pipelines or only partially agent-controlled components. The authors also release the collected training trajectories and the post-training infrastructure to enable reproduction and further research.
Who it fits and tradeoffs

Great fit if you research or build open‑world image generation systems that must integrate multi-step reasoning, web/external knowledge, and controllable generation within a single decision policy — especially for projects that can afford the engineering and compute cost of post‑training and agentic RL. Look elsewhere if you need a lightweight off‑the‑shelf T2I model for simple single-step generation: the approach adds complexity (teacher trajectory collection, RL infrastructure, reward design) and requires nontrivial compute and tooling to reproduce at scale.

Information

  • Websitearxiv.org
  • AuthorsJiahao Zhao, Xiaomin Yu, Zhongxiang Sun, Fengwei Teng, Chengwei Qin, Xiaobin Hu, Jun Xu, Shuicheng Yan
  • Published date2026/08/05

More Items

Evaluates whether MiniMax-H3 can infer latent physical events from incomplete multimodal evidence and produce videos consistent with those inferences. Builds a 517-instance benchmark across four reasoning scenarios (multi-view, audio disambiguation, video continuation, audiovisual integration) and quantifies modality-specific failure modes.

Benchmarks joint recovery of obfuscated platform messages and evidence-grounded web investigations by pairing 3,600 synthetic restoration inputs with 600 controlled web environments. Separately scores message-entry recovery and downstream web risk judgments, and provides a frozen multimodal evidence judge plus a resettable local sandbox for reproducible agent evaluation.

Measures how individual harness components—planning, action space, and context management—affect coding agents' success, cost, and behavior. Uses a modular harness across 176 matched settings on SWE-Bench Verified and Terminal-Bench 2.1 to isolate component effects and surface model- and budget-dependent trade-offs.