VEvo-Harness: Causally-Guarded Self-Evolving Harness for Agentic Image Generation

Yuan Wang; Yuhang Yu; Mengting Chen; Tianrun Xu; Zihao Pan; Xuetao Feng; Wei Li

cs.CVOct 29, 2026

VEvo-Harness: Causally-Guarded Self-Evolving Harness for Agentic Image Generation

Authors: Yuan Wang, Yuhang Yu, Mengting Chen, Tianrun Xu, Zihao Pan, Xuetao Feng, Wei Li

Organizations: Taobao & Tmall group of Alibaba · Tsinghua University

Abstract

The central bottleneck in visual generation and editing has migrated from achieving pixel-level fidelity to satisfying densely constrained user specifications through increasingly agentic workflows. Prevailing agentic generation pipelines remain shallow generate-verify-rewrite loops treating VLM judgments as coarse visual feedback, either discarding repair experience or admitting ungated updates that poison skill library. With memorize and govern periods left open, they degrade into ungoverned iterators of repeated trial-and-error rather than continual self-evolving harness. Towards this issue, we present VEvo-Harness, the first causally-guarded continual self-evolving harness system for agentic visual generation. VEvo-Harness introduces four coupled innovations: (1) a visual traceback that densifies sparse judgments into root-cause attributions with a hierarchical verification tree; (2) a visual continual governance that admits only causally valid skill patches through counterfactual dual-arm image generation replay; (3) a lifecycle-managed harness state that hierarchically distills root-cause-anchored episodes across working, experience, and pattern tiers into a persistent skill library of routable and causally verified capabilities; (4) a heterogeneous sandbox policy that orchestrates a diverse action space of repair primitives, dynamically routing cost-aware interventions based on root causes to select the most effective strategy. VEvo-Harness surpasses state-of-the-art agentic methods and rivals proprietary generators (e.g., GPT-Image, Nano-Pro) while maintaining the continual harness evolution and generalization.

Figures & tables

Figure 1: From stateless refinement to continual capability growth . VEvo-Harness causally verifies and reuses visual experience, enabling frozen generators to become progressively stronger with usage.

Figure 2: Overview of our proposed VEvo-Harness pipeline. A self-evolving scaffold is built around (a) a Interaction Loop for generation and verification, comprising (b) Visual Traceback, (c) Visual Continual Governance, (d) Lifecycle-Managed Harness State, and (e) Heterogeneous Action Sandbox.

Figure 3: Qualitative results of VEvo-Harness with baselines on generation ( left ) and editing ( right ).

Model	GenEval	GenEval2	DPG-Bench	OneIG-EN	OneIG-ZH	WISE	Avg.
Nano-Banana Comanici et al. (2025)	0.75	44.6	85.23	0.550	0.405	–	–
GPT-Image-1 Hurst et al. (2024)	0.84	–	85.15	0.533	0.474	0.80	–
Seedream-4 Seedream (2025)	0.84	–	88.25	0.576	0.553	0.78	–
Bagel Deng et al. (2025)	0.82	23.1	85.10	0.361	0.370	0.70	59.35
Ovis-U1 Wang et al. (2025)	0.89	21.7	83.72	0.382	0.395	0.64	56.02
Echo-4o Ye et al. (2025)	0.89	67.9	86.07	0.467	0.488	0.74	68.75

Table 1: Main results on general image generation benchmarks. All inference-time scaling methods are evaluated on the same frozen backends, where bold and underlined mark the best and second-best.

	LongText-EN	LongText-ZH	SpatialGenEval	CREA	ArtiMuse	Avg.
Original ( Cai et al., 2025 )	0.912 $\uparrow$ 0.000	0.932 $\uparrow$ 0.000	58.7 $\uparrow$ 0.0	11.84 $\uparrow$ 0.00	43.27 $\uparrow$ 0.00	58.41
Rewrite ( He et al., 2026b )	0.571 $\downarrow$ 0.341	0.495 $\downarrow$ 0.437	59.6 $\uparrow$ 0.9	16.75 $\uparrow$ 4.91	57.86 $\uparrow$ 14.59	56.65
Promptist ( Hao et al., 2023 )	0.034 $\downarrow$ 0.878	0.070 $\downarrow$ 0.862	48.3 $\downarrow$ 10.4	14.00 $\uparrow$ 2.16	43.74 $\uparrow$ 0.47	35.98
Search ( Ma et al., 2025 )	0.918 $\uparrow$ 0.006	0.937 $\uparrow$ 0.005	59.3 $\uparrow$ 0.6	12.98 $\uparrow$ 1.14	44.21 $\uparrow$ 0.94	59.88
Maestro ( Wan et al., 2025 )	0.877 $\downarrow$ 0.035	0.807 $\downarrow$ 0.125	60.3 $\uparrow$ 1.6	15.81 $\uparrow$ 3.97	56.86 $\uparrow$ 13.59	63.52
CRAFT ( Kovalev et al., 2025 )	0.951 $\uparrow$ 0.039	0.760 $\downarrow$ 0.172	60.6 $\uparrow$ 1.9	13.63 $\uparrow$ 1.79	54.95 $\uparrow$ 11.68	61.63

Table 2: Generalization results to downstream tasks comparison with inference-time scaling methods.

Table 3: Ablation studies of VEvo-Harness. (a) Component-level ablation analysis of Evo-Harness. (b) Micro-ablation of intra-module details, with VLM calls and wall-clock latency reported as cost.

Figure 4: Governance and Intervention Analysis. (a) Harmful admission rate under governance. (b) Per-task verification cost w/. and w/o. denied-change buffer. (c) Continuous self-evolution results comparisons on GenEval2. (d) Conditional success rate of each sandbox action per failure type.

Model	RISE-Bench				KRIS-Bench			WiseEdit	GEdit
Model	Temp.	Causal	Spatial	Logical	Factual	Concept.	Proced.	WiseEdit	GEdit
Seedream-4.0 ( Seedream, 2025 )	12.9	12.2	11.0	7.1	39.26	—	—	75.2	7.68
Nano-Banana ( Pichai, 2024 )	25.9	47.8	37.0	18.8	80.48	78.11	75.31	75.0	6.51
Nano-Banana-Pro ( Google, 2025 )	41.2	61.1	48.0	37.6	86.36	84.03	85.31	82.4	7.73
GPT-Image-1 ( Hurst et al., 2024 )	34.1	32.2	37.0	10.6	79.80	81.37	78.32	77.6	7.53
OmniGen ( Xiao et al., 2025 )	1.2	1.0	0.0	1.2	33.11	28.02	23.89	36.6	5.00

Table 4: Evaluation on instruction-driven editing benchmarks, where bold indicates the best results.

Figure 5: Generalization on different base models.

Agentic Image Generation

Agentic image generation uses reasoning, tools, search, and iterative feedback to create and refine images with greater control.

See all 45 papers

Agentic Visual Generation: From Generative Models to Agentic Control
Sep 6, 2026Yinming Huang, Shuyuan Tu, Xi Yan +7Agentic Image Generation Controllable Image Generation
Fudan University · Shanghai Innovative Institute · CUHK, MMLab +2
OmniHarness: Harnessing Generalizable Visual Generation via Symbolic Policy Learning
Sep 13, 2026Xu Xu, Jinxiu Liu, Zhangbo Qiao +5LLM Agent Harnesses Image Generation
Beihang University · The Chinese University of Hong Kong · National University of Singapore
GenEvolve: Self-Evolving Image Generation Agents via Tool-Orchestrated Visual Experience Distillation
May 20, 2026Sixiang Chen, Zhaohu Xing, Tian Ye +7Agentic Image Generation LLM Agent Self-Improvement
The Hong Kong University of Science and Technology (Guangzhou) · Meituan · The Hong Kong University of Science and Technology +1

Sep 6, 2026J/K move · Enter open · S save

Agentic Visual Generation: From Generative Models to Agentic Control

Yinming Huang, Shuyuan Tu, Xi Yan, Jiahao Zhan+6

Fudan University · Shanghai Innovative Institute · CUHK, MMLab+2

Agentic Image Generation Controllable Image Generation Agentic Video Generation Controllable Video Generation

Open paper AskPDF arXiv

Visual generation is evolving from generative models used through a single invocation into agentic control processes that can plan, select tools, inspect intermediate synthesized outputs, revise failures, and reuse prior experience. In most existing systems, the controller is an LLM or VLM, while visual generation models serve as tools or executors. However, existing work lacks a consistent criterion for determining when a generation system becomes agentic. Planning depth, tool use, multi-role collaboration, and reinforcement learning are often treated as evidence of agenticity, even though none of them necessarily determines which generation decisions the controller can make. We organize the field according to what the controller can directly control in the generation process. At L1 Conditioning Control, the controller prepares the input to a predetermined generator but does not control which visual operation is executed. At L2 Execution Control, it selects and invokes actual generation, editing, rendering, or other content-modifying operations. At L3 Outcome-Adaptive Control, it observes an intermediate outcome and uses that observation to change a subsequent operation within the current task. At L4 Experience-Adaptive Control, it retains experience from completed tasks and uses that experience to change decisions on future tasks. L0 Fixed Support separately denotes generators, editors, evaluators, reward models, benchmarks, and fixed pipelines without a deployed controller that makes generation-level decisions. These levels describe a progressively broader decision-making scope rather than model size, system complexity, output quality, tool or role count, or training method. Applying this framework across image, video, editing, 3D, world, slide, and user-interface generation reveals how controller capabilities have evolved and how their mechanisms are distributed across levels.

Abstract

Agentic Image Generation

Explore similar work

Agentic Visual Generation: From Generative Models to Agentic Control

OmniHarness: Harnessing Generalizable Visual Generation via Symbolic Policy Learning

GenEvolve: Self-Evolving Image Generation Agents via Tool-Orchestrated Visual Experience Distillation