Aditya Sanghi

Aditya Sanghi, Rao Fu, Vivian Liu et al. · stanford

Recent works have demonstrated that natural language can be used to generate and edit 3D shapes. However, these methods generate shapes with limited fidelity and diversity. We introduce CLIP-Sculptor, a method to address these constraints by producing high-fidelity and diverse 3D shapes without the need for (text, shape) pairs during training. CLIP-Sculptor achieves this in a multi-resolution approach that first generates in a low-dimensional latent space and then upscales to a higher resolution for improved shape fidelity. For improved shape diversity, we use a discrete latent space which is modeled using a transformer conditioned on CLIP's image-text embedding space. We also present a novel variant of classifier-free guidance, which improves the accuracy-diversity trade-off. Finally, we perform extensive experiments demonstrating that CLIP-Sculptor outperforms state-of-the-art baselines. The code is available at https://ivl.cs.brown.edu/#/projects/clip-sculptor.

Pradeep Kumar Jayaraman, Joseph G. Lambourne, Nishkrit Desai et al.

The Boundary representation (B-rep) format is the de-facto shape representation in computer-aided design (CAD) to model solid and sheet objects. Recent approaches to generating CAD models have focused on learning sketch-and-extrude modeling sequences that are executed by a solid modeling kernel in postprocess to recover a B-rep. In this paper we present a new approach that enables learning from and synthesizing B-reps without the need for supervision through CAD modeling sequence data. Our method SolidGen, is an autoregressive neural network that models the B-rep directly by predicting the vertices, edges, and faces using Transformer-based and pointer neural networks. Key to achieving this is our Indexed Boundary Representation that references B-rep vertices, edges and faces in a well-defined hierarchy to capture the geometric and topological relations suitable for use with machine learning. SolidGen can be easily conditioned on contexts e.g., class labels, images, and voxels thanks to its probabilistic modeling of the B-rep distribution. We demonstrate qualitatively, quantitatively, and through perceptual evaluation by human subjects that SolidGen can produce high quality, realistic CAD models.

Joseph G. Lambourne, Karl D. D. Willis, Pradeep Kumar Jayaraman et al.

Reverse Engineering a CAD shape from other representations is an important geometric processing step for many downstream applications. In this work, we introduce a novel neural network architecture to solve this challenging task and approximate a smoothed signed distance function with an editable, constrained, prismatic CAD model. During training, our method reconstructs the input geometry in the voxel space by decomposing the shape into a series of 2D profile images and 1D envelope functions. These can then be recombined in a differentiable way allowing a geometric loss function to be defined. During inference, we obtain the CAD data by first searching a database of 2D constrained sketches to find curves which approximate the profile images, then extrude them and use Boolean operations to build the final CAD model. Our method approximates the target shape more closely than other methods and outputs highly editable constrained parametric sketches which are compatible with existing CAD software.

Aditya Sanghi, Pradeep Kumar Jayaraman, Arianna Rampini et al.

Significant progress has recently been made in creative applications of large pre-trained models for downstream tasks in 3D vision, such as text-to-shape generation. This motivates our investigation of how these pre-trained models can be used effectively to generate 3D shapes from sketches, which has largely remained an open challenge due to the limited sketch-shape paired datasets and the varying level of abstraction in the sketches. We discover that conditioning a 3D generative model on the features (obtained from a frozen large pre-trained vision model) of synthetic renderings during training enables us to effectively generate 3D shapes from sketches at inference time. This suggests that the large pre-trained vision model features carry semantic signals that are resilient to domain shifts, i.e., allowing us to use only RGB renderings, but generalizing to sketches at inference time. We conduct a comprehensive set of experiments investigating different design factors and demonstrate the effectiveness of our straightforward approach for generation of multiple 3D shapes per each input sketch regardless of their level of abstraction without requiring any paired datasets during training.

Aditya Sanghi, Aliasghar Khani, Pradyumna Reddy et al.

Large-scale 3D generative models require substantial computational resources yet often fall short in capturing fine details and complex geometries at high resolutions. We attribute this limitation to the inefficiency of current representations, which lack the compactness required to model the generative models effectively. To address this, we introduce a novel approach called Wavelet Latent Diffusion, or WaLa, that encodes 3D shapes into wavelet-based, compact latent encodings. Specifically, we compress a $256^3$ signed distance field into a $12^3 \times 4$ latent grid, achieving an impressive 2427x compression ratio with minimal loss of detail. This high level of compression allows our method to efficiently train large-scale generative networks without increasing the inference time. Our models, both conditional and unconditional, contain approximately one billion parameters and successfully generate high-quality 3D shapes at $256^3$ resolution. Moreover, WaLa offers rapid inference, producing shapes within two to four seconds depending on the condition, despite the model's scale. We demonstrate state-of-the-art performance across multiple datasets, with significant improvements in generation quality, diversity, and computational efficiency. We open-source our code and, to the best of our knowledge, release the largest pretrained 3D generative models across different modalities.

Aliasghar Khani, Saeid Asgari Taghanaki, Aditya Sanghi et al.

Significant strides have been made using large vision-language models, like Stable Diffusion (SD), for a variety of downstream tasks, including image editing, image correspondence, and 3D shape generation. Inspired by these advancements, we explore leveraging these extensive vision-language models for segmenting images at any desired granularity using as few as one annotated sample by proposing SLiMe. SLiMe frames this problem as an optimization task. Specifically, given a single training image and its segmentation mask, we first extract attention maps, including our novel "weighted accumulated self-attention map" from the SD prior. Then, using the extracted attention maps, the text embeddings of Stable Diffusion are optimized such that, each of them, learn about a single segmented region from the training image. These learned embeddings then highlight the segmented region in the attention maps, which in turn can then be used to derive the segmentation map. This enables SLiMe to segment any real-world image during inference with the granularity of the segmented region in the training image, using just one example. Moreover, leveraging additional training data when available, i.e. few-shot, improves the performance of SLiMe. We carried out a knowledge-rich set of experiments examining various design factors and showed that SLiMe outperforms other existing one-shot and few-shot segmentation methods.

Ka-Hei Hui, Aditya Sanghi, Arianna Rampini et al.

Significant progress has been made in training large generative models for natural language and images. Yet, the advancement of 3D generative models is hindered by their substantial resource demands for training, along with inefficient, non-compact, and less expressive representations. This paper introduces Make-A-Shape, a new 3D generative model designed for efficient training on a vast scale, capable of utilizing 10 millions publicly-available shapes. Technical-wise, we first innovate a wavelet-tree representation to compactly encode shapes by formulating the subband coefficient filtering scheme to efficiently exploit coefficient relations. We then make the representation generatable by a diffusion model by devising the subband coefficients packing scheme to layout the representation in a low-resolution grid. Further, we derive the subband adaptive training strategy to train our model to effectively learn to generate coarse and detail wavelet coefficients. Last, we extend our framework to be controlled by additional input conditions to enable it to generate shapes from assorted modalities, e.g., single/multi-view images, point clouds, and low-resolution voxels. In our extensive set of experiments, we demonstrate various applications, such as unconditional generation, shape completion, and conditional generation on a wide range of modalities. Our approach not only surpasses the state of the art in delivering high-quality results but also efficiently generates shapes within a few seconds, often achieving this in just 2 seconds for most conditions. Our source code is available at https://github.com/AutodeskAILab/Make-a-Shape.

Qimin Chen, Johannes Merz, Aditya Sanghi et al.

We introduce UNIST, the first deep neural implicit model for general-purpose, unpaired shape-to-shape translation, in both 2D and 3D domains. Our model is built on autoencoding implicit fields, rather than point clouds which represents the state of the art. Furthermore, our translation network is trained to perform the task over a latent grid representation which combines the merits of both latent-space processing and position awareness, to not only enable drastic shape transforms but also well preserve spatial features and fine local details for natural shape translations. With the same network architecture and only dictated by the input domain pairs, our model can learn both style-preserving content alteration and content-preserving style transfer. We demonstrate the generality and quality of the translation results, and compare them to well-known baselines. Code is available at https://qiminchen.github.io/unist/.

Saeid Asgari Taghanaki, Aliasghar Khani, Ali Saheb Pasand et al.

Interpreting the learned features of vision models has posed a longstanding challenge in the field of machine learning. To address this issue, we propose a novel method that leverages the capabilities of language models to interpret the learned features of pre-trained image classifiers. Our method, called TExplain, tackles this task by training a neural network to establish a connection between the feature space of image classifiers and language models. Then, during inference, our approach generates a vast number of sentences to explain the features learned by the classifier for a given image. These sentences are then used to extract the most frequent words, providing a comprehensive understanding of the learned features and patterns within the classifier. Our method, for the first time, utilizes these frequent words corresponding to a visual representation to provide insights into the decision-making process of the independently trained classifier, enabling the detection of spurious correlations, biases, and a deeper comprehension of its behavior. To validate the effectiveness of our approach, we conduct experiments on diverse datasets, including ImageNet-9L and Waterbirds. The results demonstrate the potential of our method to enhance the interpretability and robustness of image classifiers.

Karl D. D. Willis, Pradeep Kumar Jayaraman, Hang Chu et al.

Physical products are often complex assemblies combining a multitude of 3D parts modeled in computer-aided design (CAD) software. CAD designers build up these assemblies by aligning individual parts to one another using constraints called joints. In this paper we introduce JoinABLe, a learning-based method that assembles parts together to form joints. JoinABLe uses the weak supervision available in standard parametric CAD files without the help of object class labels or human guidance. Our results show that by making network predictions over a graph representation of solid models we can outperform multiple baseline methods with an accuracy (79.53%) that approaches human performance (80%). Finally, to support future research we release the Fusion 360 Gallery assembly dataset, containing assemblies with rich information on joints, contact surfaces, holes, and the underlying assembly graph structure.

Linh Tran, Amir Hosein Khasahmadi, Aditya Sanghi et al.

Learning interpretable and human-controllable representations that uncover factors of variation in data remains an ongoing key challenge in representation learning. We investigate learning group-disentangled representations for groups of factors with weak supervision. Existing techniques to address this challenge merely constrain the approximate posterior by averaging over observations of a shared group. As a result, observations with a common set of variations are encoded to distinct latent representations, reducing their capacity to disentangle and generalize to downstream tasks. In contrast to previous works, we propose GroupVAE, a simple yet effective Kullback-Leibler (KL) divergence-based regularization across shared latent representations to enforce consistent and disentangled representations. We conduct a thorough evaluation and demonstrate that our GroupVAE significantly improves group disentanglement. Further, we demonstrate that learning group-disentangled representations improve upon downstream tasks, including fair classification and 3D shape-related tasks such as reconstruction, classification, and transfer learning, and is competitive to supervised methods.

Aditya Sanghi, Hang Chu, Joseph G. Lambourne et al.

Generating shapes using natural language can enable new ways of imagining and creating the things around us. While significant recent progress has been made in text-to-image generation, text-to-shape generation remains a challenging problem due to the unavailability of paired text and shape data at a large scale. We present a simple yet effective method for zero-shot text-to-shape generation that circumvents such data scarcity. Our proposed method, named CLIP-Forge, is based on a two-stage training process, which only depends on an unlabelled shape dataset and a pre-trained image-text network such as CLIP. Our method has the benefits of avoiding expensive inference time optimization, as well as the ability to generate multiple shapes for a given text. We not only demonstrate promising zero-shot generalization of the CLIP-Forge model qualitatively and quantitatively, but also provide extensive comparative evaluations to better understand its behavior.

Peter Meltzer, Hooman Shayani, Amir Khasahmadi et al.

Boundary Representations (B-Reps) are the industry standard in 3D Computer Aided Design/Manufacturing (CAD/CAM) and industrial design due to their fidelity in representing stylistic details. However, they have been ignored in the 3D style research. Existing 3D style metrics typically operate on meshes or pointclouds, and fail to account for end-user subjectivity by adopting fixed definitions of style, either through crowd-sourcing for style labels or hand-crafted features. We propose UVStyle-Net, a style similarity measure for B-Reps that leverages the style signals in the second order statistics of the activations in a pre-trained (unsupervised) 3D encoder, and learns their relative importance to a subjective end-user through few-shot learning. Our approach differs from all existing data-driven 3D style methods since it may be used in completely unsupervised settings, which is desirable given the lack of publicly available labelled B-Rep datasets. More importantly, the few-shot learning accounts for the inherent subjectivity associated with style. We show quantitatively that our proposed method with B-Reps is able to capture stronger style signals than alternative methods on meshes and pointclouds despite its significantly greater computational efficiency. We also show it is able to generate meaningful style gradients with respect to the input shape, and that few-shot learning with as few as two positive examples selected by an end-user is sufficient to significantly improve the style measure. Finally, we demonstrate its efficacy on a large unlabeled public dataset of CAD models. Source code and data will be released in the future.

Fenggen Yu, Zhiqin Chen, Manyi Li et al.

We introduce CAPRI-Net, a neural network for learning compact and interpretable implicit representations of 3D computer-aided design (CAD) models, in the form of adaptive primitive assemblies. Our network takes an input 3D shape that can be provided as a point cloud or voxel grids, and reconstructs it by a compact assembly of quadric surface primitives via constructive solid geometry (CSG) operations. The network is self-supervised with a reconstruction loss, leading to faithful 3D reconstructions with sharp edges and plausible CSG trees, without any ground-truth shape assemblies. While the parametric nature of CAD models does make them more predictable locally, at the shape level, there is a great deal of structural and topological variations, which present a significant generalizability challenge to state-of-the-art neural models for 3D shapes. Our network addresses this challenge by adaptive training with respect to each test shape, with which we fine-tune the network that was pre-trained on a model collection. We evaluate our learning framework on both ShapeNet and ABC, the largest and most diverse CAD dataset to date, in terms of reconstruction quality, shape edges, compactness, and interpretability, to demonstrate superiority over current alternatives suitable for neural CAD reconstruction.

Joseph G. Lambourne, Karl D. D. Willis, Pradeep Kumar Jayaraman et al.

Boundary representation (B-rep) models are the standard way 3D shapes are described in Computer-Aided Design (CAD) applications. They combine lightweight parametric curves and surfaces with topological information which connects the geometric entities to describe manifolds. In this paper we introduce BRepNet, a neural network architecture designed to operate directly on B-rep data structures, avoiding the need to approximate the model as meshes or point clouds. BRepNet defines convolutional kernels with respect to oriented coedges in the data structure. In the neighborhood of each coedge, a small collection of faces, edges and coedges can be identified and patterns in the feature vectors from these entities detected by specific learnable parameters. In addition, to encourage further deep learning research with B-reps, we publish the Fusion 360 Gallery segmentation dataset. A collection of over 35,000 B-rep models annotated with information about the modeling operations which created each face. We demonstrate that BRepNet can segment these models with higher accuracy than methods working on meshes, and point clouds.

Pradeep Kumar Jayaraman, Aditya Sanghi, Joseph G. Lambourne et al.

We introduce UV-Net, a novel neural network architecture and representation designed to operate directly on Boundary representation (B-rep) data from 3D CAD models. The B-rep format is widely used in the design, simulation and manufacturing industries to enable sophisticated and precise CAD modeling operations. However, B-rep data presents some unique challenges when used with modern machine learning due to the complexity of the data structure and its support for both continuous non-Euclidean geometric entities and discrete topological entities. In this paper, we propose a unified representation for B-rep data that exploits the U and V parameter domain of curves and surfaces to model geometry, and an adjacency graph to explicitly model topology. This leads to a unique and efficient network architecture, UV-Net, that couples image and graph convolutional neural networks in a compute and memory-efficient manner. To aid in future research we present a synthetic labelled B-rep dataset, SolidLetters, derived from human designed fonts with variations in both geometry and topology. Finally we demonstrate that UV-Net can generalize to supervised and unsupervised tasks on five datasets, while outperforming alternate 3D shape representations such as point clouds, voxels, and meshes.

Aditya Sanghi

A major endeavor of computer vision is to represent, understand and extract structure from 3D data. Towards this goal, unsupervised learning is a powerful and necessary tool. Most current unsupervised methods for 3D shape analysis use datasets that are aligned, require objects to be reconstructed and suffer from deteriorated performance on downstream tasks. To solve these issues, we propose to extend the InfoMax and contrastive learning principles on 3D shapes. We show that we can maximize the mutual information between 3D objects and their "chunks" to improve the representations in aligned datasets. Furthermore, we can achieve rotation invariance in SO${(3)}$ group by maximizing the mutual information between the 3D objects and their geometric transformed versions. Finally, we conduct several experiments such as clustering, transfer learning, shape retrieval, and achieve state of art results.

Saeid Asgari Taghanaki, Mohammad Havaei, Alex Lamb et al.

The latent variables learned by VAEs have seen considerable interest as an unsupervised way of extracting features, which can then be used for downstream tasks. There is a growing interest in the question of whether features learned on one environment will generalize across different environments. We demonstrate here that VAE latent variables often focus on some factors of variation at the expense of others - in this case we refer to the features as ``imbalanced''. Feature imbalance leads to poor generalization when the latent variables are used in an environment where the presence of features changes. Similarly, latent variables trained with imbalanced features induce the VAE to generate less diverse (i.e. biased towards dominant features) samples. To address this, we propose a regularization scheme for VAEs, which we show substantially addresses the feature imbalance problem. We also introduce a simple metric to measure the balance of features in generated images.

Aditya Sanghi, Pradeep Kumar Jayaraman

We study random embeddings produced by untrained neural set functions, and show that they are powerful representations which well capture the input features for downstream tasks such as classification, and are often linearly separable. We obtain surprising results that show that random set functions can often obtain close to or even better accuracy than fully trained models. We investigate factors that affect the representative power of such embeddings quantitatively and qualitatively.