EditHero: A Benchmark for Long-Horizon Part-Level 3D Editing and Vibe Modeling

2026-10-01

AI Summary

This study addresses the challenges of unintended degradation in non-target regions and low instruction-following fidelity during long-sequence part-level 3D editing. To this end, we construct the first benchmark for long-horizon part-level 3D editing. Methodologically, we introduce a deterministic assembly engine to generate precise targets and curate a human-verified dataset of long-sequence natural language instructions. Furthermore, we propose an LLM/VLM agent framework that integrates code verification to enable both local mesh rewriting and holistic regeneration. Experimental results demonstrate that, despite higher computational overhead, the LLM agent significantly outperforms conventional non-agent approaches in instruction-following accuracy and the preservation of unedited regions, thereby bridging a critical gap in the evaluation of iterative 3D editing.

Abstract

3D editing methods are usually tested on a single edit, yet an asset is built through a long sequence of revisions, each of which must implement the requested change while leaving everything else unchanged. We introduce EditHero, to our knowledge the first benchmark for long-horizon, part-level 3D editing, with natural-language instructions and 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. We use EditHero to compare 2 opposite approaches to 3D editing. Non-agentic methods operate top down, regenerating the object from a learned 3D representation and inferring what to keep. In contrast, LLM/VLM agents operate bottom up, editing through code that inspects the mesh and rewrites only the parts required by instructions. The non-agentic methods often miss the requested change and disturb regions that should stay fixed. Most LLMs follow instructions more closely, and all of them preserve the unedited parts better, but each of their edits takes minutes. We will release the engine and the edit sequences to support research on reliable iterative 3D editing.

Problem

• 3D editing
• long-horizon editing
• part-level editing
• benchmark
• iterative 3D editing

Innovation

• Long-horizon 3D editing
• Part-level editing benchmark
• Deterministic assembly engine
• LLM/VLM agents
• Iterative 3D editing