Vision Agent Framework OmniHarness Achieves 95% on Creative Tasks

Researchers from Beihang University, The Chinese University of Hong Kong, and the National University of Singapore have released OmniHarness, a vision agent framework that lets agents practice autonomously, inspect results, and use symbolic strategy learning to abstract verified workflows into reusable strategies.

On ComfyBench creative tasks, OmniHarness achieved a 95% solve rate, outperforming the best baseline SymbOmni by 27.5 percentage points. With the Codex GPT-4o configuration, its overall solve rate reached 92.5%. When using GPT-4o as the reasoning backbone, OmniHarness recorded an overall solve rate of 89.5% and 95% on creative tasks.

The learned strategies include workflow templates, applicability conditions, and resource dependencies, with instance-specific inputs extracted. The frozen strategy library can be adapted to other agents without fine-tuning models. In experiments, ComfyAgent's overall solve rate rose from 32.5% to 57.0%, ComfyMind from 83.0% to 88.0%, and SymbOmni from 86.0% to 89.0%.

OmniHarness also transferred an image-only strategy library to video workflows, achieving an overall video solve rate of 85.9%, 5.1 points above the best baseline, and 80.0% on video-to-video tasks, 13.3 points higher. The team noted that autonomous practice and iterative verification add computational overhead, leaving room for optimization.