Reproduce “PASTA: Controllable Part-Aware Shape Generation with Autoregressive Transformers” 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
Songlin Li, Despoina Paschalidou, Leonidas Guibas
A new model named PASTA enables controllable, part-aware generation of high-fidelity 3D shapes using an autoregressive transformer for primitive sequencing and a blending network for implicit mesh synthesis. It demonstrates superior performance in realism and diversity for shape generation and completion on ShapeNet, also facilitating conditional synthesis from text, images, or size.
Key insights
- Unifying a part-aware autoregressive primitive generator with a high-fidelity implicit blending network effectively combines controllability and geometric realism.
- Employing scheduled sampling during transformer training is crucial for mitigating exposure bias, leading to more robust and natural part arrangements during inference.
- A single, flexible model architecture can effectively perform diverse tasks, including generation from scratch, partial completion, and various forms of conditional synthesis.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.