ToolArtist Puts the Entire Image-Generation Workflow Under One Agentic Policy, Training It by Hiding the Generation Tool From Its Own SFT Traces

arXiv / HuggingFace Daily Papers·low signalXBlueskyLinkedInCopy link

arXiv 2608.04436 (August 5, Zhao, Yu, Sun et al.) post-trains a Unified Multimodal Model to coordinate reasoning, tool calls and native image synthesis under a single policy instead of a fixed pipeline. The SFT trick is unusual: they run an agent with search and image-generation tools, then reformat the data to conceal the image-generation tool while keeping the images it produced, so the model learns to emit them natively. RL uses Reason-Act-Draw GRPO with complementary intent and quality rewards; full agent control beat fixed pipelines and partial oversight, and the team committed to releasing the training data and post-training infrastructure.