Automatic Generation of Constrained Furniture Layouts

Paul Henderson, Kartic Subr, Vittorio Ferrari

Abstract

Efficiently authoring vast virtual environments requires algorithms that are both automatic and controllable. We present a fast, probabilistic generative model for indoor furniture layouts. By modeling rooms as collections of object classes and learning co-occurrence statistics from a large database (SUNCG), our method generates realistic 3D layouts through sequential conditional sampling. The system handles high-order relationships like dining sets and kitchen rows using motifs and abutments, and adds fine-grained detail via embellishments. To ensure controllability, we integrate a rejection-sampling framework that allows the model to adhere to diverse user constraints. Our results demonstrate that the generated layouts are indistinguishable from human-designed ones in quality, while being produced in milliseconds.

Introduction

The creation of detailed 3D environments for games, film, and virtual reality is a resource-intensive task. While procedural generation offers a solution, existing techniques often struggle with the balance between speed, quality, and user control. Optimization-based methods like Markov Chain Monte Carlo (MCMC) can produce high-quality results but are often too slow for interactive workflows. Conversely, many automated systems fail to capture the logical groupings of furniture—such as chairs around a table—resulting in "scattered" layouts. This work proposes a method that is simple, efficient, and highly amenable to constraints.

Methodology: Hierarchical Generative Modeling

Our approach utilizes a directed graphical model and ancestral sampling. Rather than optimizing a global energy function, the model makes a sequence of conditional decisions to populate a room:

  1. Room and Furniture Selection: The process begins by sampling a room type and then selecting furniture instances based on learned co-occurrence statistics from the SUNCG dataset.
  2. Cell-Based Layout: The room is decomposed into a cell structure (corners, edges, and interior). Objects are assigned to these cells and positioned using parametric distributions for padding and orientation.
  3. High-Order Patterns: To capture realistic groupings, we introduce "motifs" and "abutments." Motifs represent fixed groups (like a dining table and chairs), while abutments handle sequences of objects (like kitchen cabinets) modeled as a Markov chain.
  4. Embellishment: The final step adds small-scale details, such as items on a desk or ceiling fixtures, to enhance realism.
Hierarchical modeling of high-order patterns and embellishments.

Handling Constraints via Rejection Sampling

A key innovation of this work is the use of rejection sampling for constraint satisfaction. Because the unconstrained sampler is extremely fast (approximately 0.04 seconds per room), the system can generate thousands of candidate layouts in seconds. Each candidate is checked against user-defined constraints—such as room dimensions, door/window placements, or specific "must-have" items. If a layout fails, it is rejected, and a new one is sampled. This "guess and check" approach provides immense flexibility without the complexity of building constraints directly into the generative logic.

Comparison of constrained and unconstrained layout generation.

Results and Evaluation

The model was trained on the SUNCG dataset, utilizing over 45,000 human-designed rooms. To evaluate performance, we conducted a "Turing Test" style user study. Participants were presented with pairs of rooms—one human-designed and one generated by our model. Users identified the AI-generated layout as more realistic approximately 50% of the time, suggesting the outputs are indistinguishable from human work.

Furthermore, our method significantly outperformed state-of-the-art CNN-based generative models. In head-to-head comparisons, users preferred our layouts by a margin of over 70%. In terms of efficiency, while unconstrained generation is near-instantaneous, even highly constrained scenarios typically yield a valid layout in under two minutes, far faster than manual authoring.

Limitations and Future Work

We have presented a fast and controllable system for automatic furniture layout generation. Future work will address current limitations, such as the bias toward rectangular rooms and the lack of explicit aesthetic style consistency (e.g., matching furniture materials). Additionally, extending the model to support a posteriori editing—allowing users to modify a generated layout without re-sampling the entire room—remains an important research direction.

Summary updated

Abstract

Efficient authoring of vast virtual environments hinges on algorithms that are able to automatically generate content while also being controllable. We propose a method to automatically generate furniture layouts for indoor environments. Our method is simple, efficient, human-interpretable and amenable to a wide variety of constraints. We model the composition of rooms into classes of objects and learn joint (co-occurrence) statistics from a database of training layouts. We generate new layouts by performing a sequence of conditional sampling steps, exploiting the statistics learned from the database. The generated layouts are specified as 3D object models, along with their positions and orientations. We show they are of equivalent perceived quality to the training layouts, and compare favorably to a state-of-the-art method. We incorporate constraints using a general mechanism -- rejection sampling -- which provides great flexibility at the cost of extra computation. We demonstrate the versatility of our method by applying a wide variety of constraints relevant to real-world applications.
BibTeX
@misc{henderson2017automatic,
  author = {Paul Henderson and Kartic Subr and Vittorio Ferrari},
  title = {{Automatic Generation of Constrained Furniture Layouts}},
  year = {2017},
  eprint = {1711.10939},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1711.10939},
}
Figure