AutoRef: Harness Optimization for Agentic Multi-Reference Image Generation
AutoRef uses a coding agent to rewrite the image-generation harness while leaving the underlying models unchanged.
The paper targets multi-reference image generation, where models can drop, duplicate, or awkwardly paste subjects from input images. AutoRef searches over harness code, separating tasks used for feedback from tasks used to select candidates. The resulting AutoRef-Harness raises FLUX.2 4B from 5.72 to 7.37 on held-out four-reference MultiBanana tasks, matching or beating cited proprietary systems. The authors also report gains without re-optimization across other generators, reference counts, benchmarks, evaluators, and reasoning models. HF Daily Papers' note
The paper targets multi-reference image generation, where models can drop, duplicate, or awkwardly paste subjects from input images. AutoRef searches over harness code, separating tasks used for feedback from tasks used to select candidates. The resulting AutoRef-Harness raises FLUX.2 4B from 5.72 to 7.37 on held-out four-reference MultiBanana tasks, matching or beating cited proprietary systems. The authors also report gains without re-optimization across other generators, reference counts, benchmarks, evaluators, and reasoning models. HF Daily Papers' note
score 4