[Paper Review] House-GAN++: Generative Adversarial Layout Refinement Networks

Nelson Nauata, Sepidehsadat Hosseini+4 more|arXiv (Cornell University)|Mar 3, 2021

Generative Adversarial Networks and Image SynthesisComputer Science30 references19 citations

TL;DR

This paper proposes House-GAN++, a generative adversarial network that iteratively refines floorplan layouts using a graph-constrained relational GAN and conditional GAN architecture, where each refinement step uses the previous output as a constraint. The key contribution is a non-iterative training method—component-wise GT-conditioning—that effectively trains an iterative generator, enabling the system to produce layouts nearly indistinguishable from professional architect-designed floorplans, even outperforming state-of-the-art methods on realism, diversity, and compatibility metrics.

ABSTRACT

This paper proposes a novel generative adversarial layout refinement network for automated floorplan generation. Our architecture is an integration of a graph-constrained relational GAN and a conditional GAN, where a previously generated layout becomes the next input constraint, enabling iterative refinement. A surprising discovery of our research is that a simple non-iterative training process, dubbed component-wise GT-conditioning, is effective in learning such a generator. The iterative generator also creates a new opportunity in further improving a metric of choice via meta-optimization techniques by controlling when to pass which input constraints during iterative layout refinement. Our qualitative and quantitative evaluation based on the three standard metrics demonstrate that the proposed system makes significant improvements over the current state-of-the-art, even competitive against the ground-truth floorplans, designed by professional architects.

Motivation & Objective

• To address the challenge of automated, high-quality floorplan generation that mimics the iterative design process of professional architects.
• To enable iterative layout refinement in a GAN-based framework despite the lack of real incomplete floorplan datasets.
• To develop a training method that effectively learns an iterative generator without requiring iterative training data.
• To introduce test-time meta-optimization for further improving specific metrics like compatibility or diversity.
• To evaluate the system against professional architects and demonstrate near-human-level performance in layout quality.

Proposed method

• The architecture integrates a graph-constrained relational GAN with a conditional GAN, where each refinement step uses the prior layout as a conditional input constraint.
• A novel non-iterative training method, component-wise GT-conditioning, passes ground-truth segmentation masks to each component at random probabilities to train the iterative generator.
• The generator produces vector floorplans as a sequence of refinement steps, improving spatial arrangement, room shapes, and door placement over iterations.
• Test-time meta-optimization controls when to pass which input constraints (e.g., room masks) to optimize for specific metrics like compatibility or diversity.
• The system uses the RPLAN dataset of 60k professional floorplans and corresponding bubble diagrams for training and evaluation.
• Evaluation uses standard metrics: realism (via user study), diversity, and compatibility, with pairwise comparisons against SOTA and ground-truth.

Experimental results

Research questions

• RQ1 Can a non-iterative training process effectively train a generator that performs iterative layout refinement?
• RQ2 Can iterative refinement via a GAN-based system produce floorplans that are indistinguishable from professional architect designs?
• RQ3 How effective is component-wise GT-conditioning in enabling the learning of an iterative refinement generator?
• RQ4 Can test-time meta-optimization improve specific layout metrics such as compatibility or diversity?
• RQ5 To what extent can the generated layouts match or exceed the quality of ground-truth floorplans in realism and spatial coherence?

Key findings

• The proposed system, House-GAN++, outperforms all competing methods on all three standard metrics—realism, diversity, and compatibility—by large margins.
• In user studies, professional architects could not reliably distinguish the best variant of House-GAN++ (Ours_static*) from ground-truth floorplans, with 25% choosing it as better and 32% calling it a tie.
• The system achieved a pairwise realism score of -0.18 against ground-truth in vector-floorplan visualization, indicating near-indistinguishability.
• Meta-optimization significantly improved metric-specific performance: compatibility was maximized when optimized for compatibility, and diversity when optimized for diversity.
• The optimal refinement scheme passed mask-conditions earlier for compatible layouts than incompatible ones, and for doors before larger rooms, contrary to intuitive design order.
• Even with only 10% of samples being fully compatible, the refined system (Ours_static*) achieved performance close to ground-truth, demonstrating strong practical viability.