AutoRef: Harness Optimization for Agentic Multi-Reference Image Generation
AutoRef automatically rewrites image-generation harnesses, lifting FLUX.2 klein 4B from 5.72 to 7.37 on MultiBanana.
AutoRef automatically optimizes the executable harness for agentic multi-reference image generation while keeping the generator and reasoning models frozen. A coding agent iteratively rewrites harness code, using separate tasks for proposing changes and selecting candidates and continuing from a beam of top-ranked harnesses. The resulting AutoRef-Harness raised open-weight FLUX.2 klein 4B from 5.72 to 7.37 on held-out four-reference MultiBanana tasks, matching or exceeding Nano Banana Pro and GPT-Image-1.5. Without re-optimization, the same harness also improved results when the generator, reference count, benchmark, evaluator, or reasoning model changed.