Agent Banana: High-Fidelity Image Editing with Agentic Thinking and Tooling

Ruijie Ye, Jiayi Zhang, Zhuoxin Liu, Zihao Zhu, Siyuan Yang, Li Li, Tianfu Fu, Franck Dernoncourt, Yue Zhao, Jiacheng Zhu, Ryan Rossi, Wenhao Chai, Zhengzhong Tu

Abstract

We study instruction-based image editing under professional workflows and identify three persistent challenges: (i) editors often over-edit, modifying content beyond the user's intent; (ii) existing models are largely single-turn, while multi-turn edits can alter object faithfulness; and (iii) evaluation at around 1K resolution is misaligned with real workflows that often operate on ultra high-definition images (e.g., 4K). We propose Agent Banana, a hierarchical agentic planner-executor framework for high-fidelity, object-aware, deliberative editing. Agent Banana introduces two key mechanisms: (1) Context Folding, which compresses long interaction histories into structured memory for stable long-horizon control; and (2) Image Layer Decomposition, which performs localized layer-based edits to preserve non-target regions while enabling native-resolution outputs. To support rigorous evaluation, we build HDD-Bench, a high-definition, dialogue-based benchmark featuring verifiable stepwise targets and native 4K images (11.8M pixels) for diagnosing long-horizon failures. On HDD-Bench, Agent Banana achieves the best multi-turn consistency and background fidelity (e.g., IC 0.871, SSIM-OM 0.84, LPIPS-OM 0.12) while remaining competitive on instruction following, and also attains strong performance on standard single-turn editing benchmarks. We hope this work advances reliable, professional-grade agentic image editing and its integration into real workflows.

Introduction

Instruction-based image editing has evolved rapidly, yet a significant gap remains between current generative models and the requirements of professional creative workflows. While models can easily perform atomic edits like changing a car's color, they often struggle with the nuances of iterative, high-resolution editing. We identify three critical limitations in existing systems: over-editing, where non-target regions are unintentionally modified; multi-turn degradation, where sequential edits introduce accumulating artifacts; and resolution constraints, which typically downsample images to 1K, sacrificing the fine detail necessary for 4K+ professional production.

The Agent Banana Framework

To overcome these hurdles, we introduce Agent Banana, a hierarchical multi-agent framework designed for high-fidelity, multi-turn editing. At its core, the system decouples global reasoning from low-level execution. A Planner agent (powered by GPT-5-mini) interprets user instructions and decomposes them into a series of sub-goals. An Executor agent then manages the specific tool invocations—such as adding, removing, or adjusting elements—and performs intermediate validations to ensure each step aligns with the user's intent.

Context Folding and Image Layer Decomposition

Maintaining consistency over long dialogues is a major challenge for LLM-based agents. Agent Banana employs Context Folding, a memory management technique that organizes history into three distinct tiers:

  1. Asset Level: Semantic text nodes that track the state of various image elements.
  2. Execution Level: Transient memory used for trial-and-error during tool invocation, which is discarded once a step is finalized.
  3. Planning Level: A persistent record of verified, successful edits.

This separation prevents the Planner from becoming overwhelmed by the "noise" of intermediate execution failures. To address the resolution and over-editing problems, we implement Image Layer Decomposition (ILD). Instead of processing the entire image through a diffusion model, ILD identifies target regions, crops them into high-resolution patches, and performs edits locally. These patches are then blended back into the original 4K canvas. This approach mathematically guarantees that unedited regions remain untouched, preventing Prior-Induced Editing Drift (PIED)—a phenomenon where background textures slowly drift toward model-specific statistics over multiple turns.

HDD-Bench: A New Standard for Evaluation

Rigorous evaluation of multi-turn, high-definition editing requires new benchmarks. We present HDD-Bench, a High-Definition, Dialogue-based benchmark featuring 96 curated 3-turn editing sessions at native 4K resolution (11.8M pixels). Unlike previous benchmarks, HDD-Bench utilizes a scalable symbolic data engine to track object attributes (e.g., name, color, size) across turns. This enables verifiable, step-by-step evaluation of an agent's performance without the need for expensive human-in-the-loop annotation.

Results and Performance

Agent Banana sets a new state-of-the-art on HDD-Bench, particularly in multi-turn consistency and background preservation. It achieves an Image Consistency (IC) score of 0.871 and an instruction-following score of 0.849. In terms of fidelity, the system vastly outperforms baselines, reaching an Otsu-Masked SSIM of 0.84 and an LPIPS of 0.12. These metrics specifically measure the preservation of non-target regions, highlighting the effectiveness of the ILD mechanism. Furthermore, the system demonstrates strong performance on single-turn benchmarks like ImgEdit-Bench, proving that the agentic overhead does not compromise basic editing precision.

Conclusion

Agent Banana demonstrates that combining agentic reasoning with localized, layer-based execution is a viable path toward professional-grade image editing. By addressing the issues of over-editing, multi-turn degradation, and resolution limits, this framework provides a more reliable tool for complex creative tasks. Future research will focus on reducing the dependency on specific foundational LLMs and implementing robust provenance mechanisms to ensure the ethical use of high-fidelity editing technologies.

Summary updated

Abstract

We study instruction-based image editing under professional workflows and identify three persistent challenges: (i) editors often over-edit, modifying content beyond the user's intent; (ii) existing models are largely single-turn, while multi-turn edits can alter object faithfulness; and (iii) evaluation at around 1K resolution is misaligned with real workflows that often operate on ultra high-definition images (e.g., 4K). We propose Agent Banana, a hierarchical agentic planner-executor framework for high-fidelity, object-aware, deliberative editing. Agent Banana introduces two key mechanisms: (1) Context Folding, which compresses long interaction histories into structured memory for stable long-horizon control; and (2) Image Layer Decomposition, which performs localized layer-based edits to preserve non-target regions while enabling native-resolution outputs. To support rigorous evaluation, we build HDD-Bench, a high-definition, dialogue-based benchmark featuring verifiable stepwise targets and native 4K images (11.8M pixels) for diagnosing long-horizon failures. On HDD-Bench, Agent Banana achieves the best multi-turn consistency and background fidelity (e.g., IC 0.871, SSIM-OM 0.84, LPIPS-OM 0.12) while remaining competitive on instruction following, and also attains strong performance on standard single-turn editing benchmarks. We hope this work advances reliable, professional-grade agentic image editing and its integration into real workflows.
BibTeX
@misc{ye2026agent,
  author = {Ruijie Ye and Jiayi Zhang and Zhuoxin Liu and Zihao Zhu and Siyuan Yang and Li Li and Tianfu Fu and Franck Dernoncourt and Yue Zhao and Jiacheng Zhu and Ryan Rossi and Wenhao Chai and Zhengzhong Tu},
  title = {{Agent Banana: High-Fidelity Image Editing with Agentic Thinking and Tooling}},
  year = {2026},
  eprint = {2602.09084},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2602.09084},
}
Figure