Researchers have introduced EditHero, described as the first benchmark for long-horizon, part-level 3D editing. The work addresses a gap in how 3D editing methods are evaluated: while real assets are built through long sequences of revisions, most methods are tested on a single edit.
EditHero provides natural-language instructions and target images covering both geometry and texture. A deterministic assembly engine produces the exact target after every edit, and each sequence is reviewed by hand, according to the paper.
The benchmark was used to compare two opposing approaches. Non-agentic methods work top down, regenerating the object from a learned 3D representation and inferring what to keep. LLM/VLM agents work bottom up, editing through code that inspects the mesh and rewrites only the parts required by instructions.
According to the findings, 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 unedited parts better, but each edit takes minutes.
The authors plan to release the engine and the edit sequences to support further research on reliable iterative 3D editing.