EvoGen-Harness: Learning Where and How to Evolve Image-Generation Harnesses

2026-09-30

AI Summary

This study addresses the limitation of existing text-to-image systems, where single-dimensional optimization constrains the adaptation space, by proposing a generator-agnostic multi-responsibility evolutionary framework. The core innovation lies in introducing a Trace mechanism that leverages failure attribution as a search prior, coordinating the joint evolution of external components through trajectory-based attribution. Furthermore, strategies including stochastic execution evidence aggregation, No-Patch verification, and hold-out set validation are designed to ensure optimization reliability. Experimental results demonstrate that the proposed method significantly outperforms baselines across multiple benchmarks, achieving an attribution recall rate of 87.9% with a regression rate of only 1.9%.

Abstract

Modern text-to-image (T2I) systems can be improved without modifying generator parameters by adapting the external system around frozen generators. However, existing approaches typically optimize a predefined dimension, such as prompts, routing, or workflows, restricting the space in which generation failures can be corrected. Allowing multiple generator-external responsibilities to evolve provides a broader adaptation space, but introduces a new challenge: visual feedback reveals what failed, but not where persistent evolution should occur or how this space should be explored efficiently. We introduce EvoGen-Harness, a generator-agnostic framework for multi-responsibility image-generation harness evolution, together with Trace (Trajectory-Relative Attribution and Coordinated Evolution). Trace aggregates evidence across stochastic executions, uses failure attribution as a search prior to focus candidate updates, and progressively re-attributes residual failures to coordinate evolution across responsibilities, while No-Patch and held-out validation prevent unnecessary or harmful updates. Across GenEval2, T2I-CompBench++, and WISE, EvoGen-Harness improves over the strongest evaluated baselines by +0.2633, +0.0720, and +0.0752, respectively, while achieving 87.9-91.4% attribution recall, 94.8% No-Patch accuracy, and only 1.9% regression. These results demonstrate that attribution-guided multi-responsibility evolution can substantially enhance frozen T2I systems beyond single-dimension adaptation.

Problem

• Text-to-Image Generation
• Multi-responsibility Evolution
• Failure Attribution
• Frozen Generators
• External Adaptation

Innovation

• Multi-responsibility evolution
• Failure attribution
• Generator-agnostic framework
• Text-to-image generation
• Trace algorithm