ToolArtist: Tool-Using Unified Multimodal Models for Agentic Image Generation
ToolArtist puts reasoning, tool calls, and image generation under one agent policy.
The paper describes a unified multimodal model post-trained to handle open-world image-generation tasks with dynamic reasoning and external tool use. Its SFT stage uses a teacher agent with search and image-generation tools, then hides the image-generation tool while keeping the resulting images for training. Its RL stage introduces RAD-GRPO, with intent and quality rewards for the model’s Reason-Act-Draw loop. The authors say this full agent-control setup beats fixed pipelines and partially agentic approaches, and they release the training data and post-training infrastructure. HF Daily Papers' note
The paper describes a unified multimodal model post-trained to handle open-world image-generation tasks with dynamic reasoning and external tool use. Its SFT stage uses a teacher agent with search and image-generation tools, then hides the image-generation tool while keeping the resulting images for training. Its RL stage introduces RAD-GRPO, with intent and quality rewards for the model’s Reason-Act-Draw loop. The authors say this full agent-control setup beats fixed pipelines and partially agentic approaches, and they release the training data and post-training infrastructure. HF Daily Papers' note
score 5