Overview
3D editing methods are typically evaluated on a single modification. In practice, however, a 3D asset is built through a long sequence of revisions, and each revision must implement the requested change while leaving everything else untouched.
To address this gap, researchers introduce EditHero, described as the first benchmark for long-horizon, part-level 3D editing. It provides natural-language instructions along with target images for both geometry and texture. A deterministic assembly engine produces the exact target after every edit, and every sequence is reviewed by hand to ensure correctness.
Benchmark Design
EditHero is designed to test whether a system can follow an extended chain of editing instructions without drifting from the intended result. Each edit sequence is verified manually, and the assembly engine guarantees a precise ground-truth target at every step. This setup makes it possible to measure both instruction adherence and preservation of unedited regions over many successive operations.
Comparing Two Approaches to 3D Editing
The benchmark is used to compare two opposite approaches:
- Non-agentic methods: These operate top down, regenerating the object from a learned 3D representation and inferring which parts should be kept.
- LLM/VLM agents: These operate bottom up, editing through code that inspects the mesh and rewrites only the parts required by the instructions.
Findings
According to the results, non-agentic methods often miss the requested change and disturb regions that should remain fixed. Most LLMs follow instructions more closely, and all of them preserve the unedited parts better. However, each of their edits takes minutes, highlighting a trade-off between reliability and speed.
Availability
The project page and code are available online. The authors plan to release the engine and the edit sequences to support research on reliable iterative 3D editing.
2026 Context
As of 2026, long-horizon 3D editing remains a challenging problem, particularly when edits must be applied at the part level and preserve the rest of the asset. Benchmarks like EditHero are increasingly important as LLM and VLM agents become more capable of code-based mesh manipulation, while non-agentic generative methods continue to struggle with precise, iterative changes. The release of the engine and sequences aims to accelerate progress toward more dependable iterative 3D editing systems.
Paper Details
- Title: EditHero: A Benchmark for Long-Horizon Part-Level 3D Editing and Vibe Modeling
- Authors: Ruihan Yu, Yu-Ju Tsai, Muyao Niu, Runyi Li, Lian Fu, Hanqing Liu, Zheng-Hui Huang, Yonghao Yu, Sho Kuno, Ming-Hsuan Yang, Kaipeng Zhang, Zhixiang Wang
- Submitted: 1 Oct 2026
- Subjects: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI); Graphics (cs.GR)
- arXiv: 2610.02298
- Project page: https://alaya-lab.github.io/EditHero/
- Code: https://github.com/AlayaLab/EditHero
via ArXiv CV