OmniHarness: Harnessing Generalizable Visual Generation via Symbolic Policy Learning

OmniHarness is a visual-generation harness that turns verified ComfyUI executions into reusable symbolic policies: instance-specific prompts and images are removed, while procedures, preconditions, dependencies, and failure remedies remain available for retrieval, adaptation, and composition. Before receiving downstream objectives, it performs self-directed inquiry, selecting practice tasks with capability novelty and a competence-frontier heuristic; during execution, ordered planning, intermediate verification, and localized recovery prevent errors from propagating, while policy updates occur with model weights frozen. After 50 inquiry iterations, the system learned 43 symbolic policies and was evaluated across ComfyBench, GenEval, GenEval2, Reason-Edit, WISE, and KRIS-Bench using multiple agent frameworks and models including OpenAI’s GPT-4o, Google’s Gemini-2.5-Flash, DeepSeek-V3, and GPT-Image-1. On ComfyBench Creative tasks, Codex GPT-4o with OmniHarness achieved a 95.0% Resolve rate, 27.5 percentage points above the strongest baseline; overall Resolve reached 92.5%. It scored 0.997 on GenEval and 0.86 on WISE, while image-only policies transferred to unseen video tasks with an 85.9% overall Resolve rate. Removing self-directed inquiry reduced Creative Resolve from 95.0% to 72.5%, and disabling online policy updates reduced overall Resolve to 88.5%, supporting both proactive skill acquisition and continual feedback-driven refinement. A frozen policy snapshot also improved ComfyAgent, ComfyMind, and SymbOmni without fine-tuning, indicating portability across visual-agent architectures and modalities.