Reproduce “TexEditor: Structure-Preserving Text-Driven Texture Editing” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Bo Zhao, Yihang Liu, Chenfeng Zhang, Huan Yang, Kun Gai, Wei Ji

TexEditor introduces a two-stage training framework for text-driven texture editing that robustly preserves object geometric structure. The model consistently modifies object appearance according to text prompts while maintaining fidelity to underlying shapes, achieving superior performance on the new real-world `TexBench` benchmark with a `TexEval` score of 0.886 for texture replacement.

Key insights

  • Training with a dedicated, complex scene-level synthetic dataset (TexBlender) is crucial for a cold start that instills strong structural priors in texture editing models.
  • Reinforcement learning, guided by a reward function that explicitly enforces structural consistency alongside instruction adherence, effectively transfers learned structural priors to real-world scenarios.
  • A new comprehensive evaluation framework, combining the real-world `TexBench` benchmark and the `TexEval` metric (which balances MLLM-based instruction adherence with SAUGE-based structural similarity), provides a more robust assessment of texture editing quality.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.