ShapeCrafter: A Recursive Text-Conditioned 3D Shape Generation Model

Authors: Rao Fu, Xiao Zhan, YIWEN CHEN, Daniel Ritchie, Srinath Sridhar

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments and results produced by Shape Crafter demonstrate that we can generate and evolve high-quality 3D shape distributions that are consistent with text inputs. In this section, we provide both quantitative and qualitative evaluation of Shape Crafter on recursive text-conditioned shape generation task.
Researcher Affiliation Academia Brown University rao_fu@brown.edu ivl.brown.edu/projects/shapecrafter
Pseudocode No The paper does not contain a dedicated pseudocode or algorithm block.
Open Source Code No The paper does not provide an explicit statement about releasing source code or a link to a code repository.
Open Datasets Yes Therefore, we construct a new dataset, Text2Shape++, building upon the Text2Shape [6] dataset, which has text shape pairs of chair and table categories from Shape Net [5]. Text2Shape++ contains 369K text shape pairs, which to our knowledge is the largest dataset of its kind.
Dataset Splits Yes All our experiments use our Text2Shape++ dataset for training (20% held out for validation).
Hardware Specification No The paper does not provide specific hardware details (e.g., exact GPU/CPU models, memory amounts, or detailed computer specifications) used for running its experiments.
Software Dependencies No The paper does not provide specific software dependencies with version numbers.
Experiment Setup Yes In our method, we empirically set K = 512, D = 256, and the 3D shape grid resolution is g = 8. For each input text I, we use the first token in last layer of the BERTBASE model as the text feature embedding B R768. Our loss function consists of reconstruction loss, vector quantization objective, and commitment loss as proposed by van den Oord et al. [47]