cs.LGSep 30, 2026
SaveEvoGen-Harness: Learning Where and How to Evolve Image-Generation Harnesses
Organizations: Peking University · Beijing University of Technology · Fudan University
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.
Figures & tables
Figure 2: Motivation. Prior image-generation methods typically optimize a predefined adaptation dimension, restricting the available correction space. Allowing multiple harness responsibilities to evolve provides a broader adaptation space, but introduces a new attribution problem: the same visual failure may require different persistent changes, or no update at all.
| Type | Method | Object | Attribute | Count | Position | Verb | Overall |
|---|---|---|---|---|---|---|---|
| T2I Models | FLUX.1-dev | 0.8754 | 0.6734 | 0.5311 | 0.3928 | 0.2150 | 0.2115 |
| Janus-Pro | 0.9102 | 0.7213 | 0.4834 | 0.5217 | 0.2056 | 0.1878 | |
| Flow-GRPO | 0.9175 | 0.7438 | 0.6032 | 0.5233 | 0.2541 | 0.2198 | |
| Qwen-Image | 0.9730 | 0.8412 | 0.6857 | 0.6083 | 0.3752 | 0.3483 | |
| Z-Image | 0.9755 | 0.7851 | 0.6232 | 0.6041 | 0.2419 | 0.3034 | |
| Nano Banana | 0.9722 | 0.8981 | 0.6830 | 0.7044 | 0.4532 | 0.4456 |
Table 1: Quantitative evaluation on the GenEval2 benchmark. We report atom-level Soft-TIFA AM across five visual skills and prompt-level Soft-TIFA GM as the overall score. The best and second-best results are marked in bold red and underlined blue , respectively. indicates higher is better.
| Method | Color | Shape | Texture | 2D-Spatial | 3D-Spatial | Numeracy | Non-Spatial | Complex | Average |
|---|---|---|---|---|---|---|---|---|---|
| T2I Models | |||||||||
| FLUX.1-dev | 0.7572 | 0.5066 | 0.6300 | 0.2700 | 0.3992 | 0.6165 | 0.3065 | 0.3628 | 0.4811 |
| Janus-Pro | 0.5145 | 0.3323 | 0.4069 | 0.1566 | 0.2753 | 0.4406 | 0.3137 | 0.3806 | 0.3526 |
| Flow-GRPO | 0.8379 | 0.6130 | 0.7236 | 0.5447 | 0.4471 | 0.6752 | 0.3195 | 0.3741 | 0.5669 |
| Qwen-Image | 0.8392 | 0.5941 | 0.7568 | 0.4512 | 0.4631 | 0.7705 | 0.3126 | 0.3944 | 0.5727 |
| Z-Image | 0.8413 | 0.6078 | 0.7512 | 0.3822 | 0.4412 | 0.6676 | 0.3181 | 0.4201 | 0.5537 |
Table 2: Quantitative evaluation results on T2I-CompBench++. The best and second-best results are marked in bold red and underlined blue , respectively. indicates higher is better.
| Method | Culture | Time | Space | Biology | Physics | Chemistry | Overall |
|---|---|---|---|---|---|---|---|
| Qwen-Image | 0.6275 | 0.5250 | 0.5583 | 0.3417 | 0.4833 | 0.2500 | 0.5100 |
| Nano Banana | 0.7000 | 0.5900 | 0.6700 | 0.4500 | 0.5800 | 0.4000 | 0.6028 |
| VisualPrompter | 0.5800 | 0.4500 | 0.5900 | 0.2300 | 0.4300 | 0.2900 | 0.4708 |
| OctoT2I | 0.5350 | 0.4160 | 0.5090 | 0.2620 | 0.5130 | 0.3140 | 0.4557 |
| Ours | 0.7589 | 0.6712 | 0.7160 | 0.5889 | 0.6651 | 0.4788 | 0.6780 |
Table 3: Quantitative evaluation on the WISE benchmark. We report world-knowledge consistency across six domains and the overall score. The best and second-best results are marked in bold red and underlined blue , respectively. indicates higher is better.
| Method | Object | Attribute | Count | Position | Verb | Overall |
|---|---|---|---|---|---|---|
| FLUX.1-dev | 0.8754 | 0.6734 | 0.5311 | 0.3928 | 0.2150 | 0.2115 |
| + Promptist | 0.8533 | 0.6510 | 0.5122 | 0.3819 | 0.2134 | 0.2056 |
| + VisualPrompter | 0.9321 | 0.7223 | 0.5614 | 0.4350 | 0.2412 | 0.2552 |
| + Ours | 0.9799 | 0.9469 | 0.9053 | 0.6418 | 0.3851 | 0.7089 |
| over VisualPrompter | +0.0478 | +0.2246 | +0.3439 | +0.2068 | +0.1439 | +0.4537 |
| Qwen-Image | 0.9730 | 0.8412 | 0.6857 | 0.6083 | 0.3752 | 0.3483 |
Table 4: Generalization across frozen T2I backbones on GenEval2. We apply the same generator-external adaptation methods to different frozen T2I backbones and we report Soft-TIFA AM for skill-wise analysis and Soft-TIFA GM for overall performance. denotes the absolute gain of EvoGen-Harness over baseline on the corresponding backbone. The best results within each backbone are marked in bold . indicates higher is better.
Figure 3: Harness attribution across persistent failure causes.
Figure 4: Progressive visual repair with Trace . Trace progressively repairs multiple visual constraints through successive harness updates while keeping the generator frozen.
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
| Responsibility | Persistent representation | Editable contents |
|---|---|---|
| Policy | Visual requirement and constraint rules | Requirement interpretation, constraint priority, and criteria used to represent explicit visual requirements such as count, attribute, relation, and action constraints. |
| Tools | Capability registry for available generation and auxiliary tools | Capability and limitation descriptions, applicable scenarios, and tool-specific knowledge used by the harness. Tool implementations and model parameters are never modified. |
| Skills | Reusable procedural templates | Multi-step generation routines, prompt templates, decomposition strategies, and reusable task-solving procedures. |
| Middleware | Run-time orchestration configuration | Routing, verification, retry, termination, and tool-invocation rules. The deterministic base controller itself remains fixed. |
| Memory | Persistent cross-task experience records | Reusable summaries of successful and failed executions, recurring failure patterns, and previously validated corrective experience. |
Table 8: Operational representation of the five persistent harness responsibilities. The examples describe the editable state rather than the parameters of the frozen generator.
| Search Strategy | GenEval2 | Candidate Edits | Gen. Calls | Evol. Time (s) |
|---|---|---|---|---|
| Full-Space Search | 0.7112 | 20.0 | 80.0 | 2386 |
| Hard Top-1 | 0.6815 | 4.0 | 16.0 | 486 |
| Trace Top- | 0.7089 | 8.0 | 32.0 | 964 |
Table 9: Search efficiency of Trace . Top- localization substantially reduces exploration relative to full-space search while preserving generation quality. Candidate edits and generator calls are averaged per task. / indicate higher / lower is better.
| Compound Causes | Cause Recall@3 | Both Recovered | Repair Success |
|---|---|---|---|
| Policy + Tools | 96.1% | 92.4% | 86.8% |
| Policy + Skills | 94.8% | 90.1% | 85.7% |
| Policy + Middleware | 94.6% | 89.5% | 84.3% |
| Policy + Memory | 95.5% | 91.3% | 85.1% |
| Tools + Skills | 94.1% | 88.7% | 83.6% |
| Tools + Middleware | 93.8% | 87.9% | 82.4% |
Table 10: Robustness of Trace under paired-responsibility failures. Each example contains two simultaneously injected persistent causes. Cause Recall@3 measures coverage of the two ground-truth causes by the initial Top-3 localization frontier; Both Recovered measures whether both causes are identified over progressive re-localization; and Repair Success measures whether both injected failure effects are resolved after evolution. indicates higher is better.
Figure 5: Sensitivity analysis of Trace . (a) A larger attribution frontier initially improves GenEval2 performance by retaining more plausible responsibilities, but yields diminishing gains while increasing the number of candidate edits. (b) Increasing improves attribution reliability by reducing the influence of stochastic generation, with gains saturating as generator calls increase. (c) Increasing improves held-out performance by enabling residual failures to be re-attributed across responsibilities, while deeper evolution provides limited additional benefit at higher cost.
| Proposal LLM | GenEval2 | Held-out Gain | Commit Rate | Regression |
|---|---|---|---|---|
| Qwen3.5-27B | 0.6996 | +0.133 | 70.6% | 2.3% |
| GPT-5.5 | 0.7164 | +0.146 | 73.2% | 1.7% |
| GPT-4.1 (Default) | 0.7089 | +0.140 | 72.0% | 1.9% |
Table 11: Robustness to different LLMs on GenEval2. Only the proposal LLM is replaced while all other components and budgets remain fixed.
Figure 6: Responsibility-specific harness evolution. Representative Policy, Skills, Tools, and Memory updates repair distinct persistent failure modes while the image generator remains frozen.
Figure 7: Cross-generator adaptation. Representative qualitative results of EVOGEN-HARNESS on FLUX.1-dev, Qwen-Image, and Janus-Pro. Each image generator remains frozen while its generator-external harness is evolved.