Reproduce “Prox-E: Fine-Grained 3D Shape Editing via Primitive-Based Abstractions” 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
Etai Sella, Hao Phung, Nitay Amiel, Or Litany, Or Patashnik, Hadar Averbuch-Elor
Prox·E is a training-free framework that enables fine-grained 3D shape editing from text by leveraging primitive-based shape abstractions and a Vision-Language Model. It achieved superior results in balancing identity preservation, 3D shape quality, and fidelity to editing prompts, with an LPIPS score of 0.10 for identity preservation and a VQAScore of 0.71 for edit fidelity.
Key insights
- Primitive-based shape abstractions provide an explicit, structured 3D vocabulary that enables VLMs to perform precise spatial reasoning for executing fine-grained structural edits from text.
- An iterative visual verification loop, feeding renderings of edited proxies back to the VLM, significantly enhances the robustness and accuracy of primitive-level modifications.
- Blending specific latent features from the original and a warped shape, alongside the edited proxy, within the denoising process is critical for seamlessly integrating changes while preserving the integrity of unedited regions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.