EditHero: A Benchmark for Long-Horizon Part-Level 3D Editing and Vibe Modeling
EditHero tests whether 3D editors can survive a chain of precise part-level changes without damaging the rest of the asset.
The benchmark pairs natural-language edit instructions with target images for both geometry and texture. Its targets are produced by a deterministic assembly engine, then hand-reviewed. In the paper’s comparison, top-down non-agentic methods often miss edits and disturb unchanged regions, while LLM/VLM agents preserve untouched parts better but take minutes per edit. HF Daily Papers' note
The benchmark pairs natural-language edit instructions with target images for both geometry and texture. Its targets are produced by a deterministic assembly engine, then hand-reviewed. In the paper’s comparison, top-down non-agentic methods often miss edits and disturb unchanged regions, while LLM/VLM agents preserve untouched parts better but take minutes per edit. HF Daily Papers' note
score 4