Research map / Image generation / Autoregressive and latent generation
MaskPLAN: Masked Generative Layout Planning from Partial Input
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes MaskPLAN, a graph-structured dynamic masked autoencoder with five attribute-specific transformer generators that performs user-guided floorplan layout generation from partial attribute inputs.
Problem
Floorplan layout generation tools cannot accept partial, incomplete design inputs, do not support flexible user-specified attributes, and fail to model interdependencies between functional and compositional layout attributes.
Approach
Frame layout generation as predicting unobserved attributes from a masked attribute matrix using dynamic masking (50%-100% uniform), a partial-input encoder used as a global prior, cross-attention conditioning across sequentially decoded attribute generators (graph then shapes), and visual tokens from a pretrained VQ-VAE (ADLM).
Key results
- Our III (graph input given) achieves fid_img 0.139 and mseT 0.00001, outperforming all baselines including iPLAN* (0.241) and Graph2Plan (1.290).
- Our I, conditioned only on the boundary, reaches fid_img 4.182, close to iPLAN (3.192) and Graph2Plan performance levels.
- Dynamic 50%-100% masking gives the best fid_img of 4.182 versus 6.478 for 0%-100% and 5.169 for 25%-100% in ablation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Condition on room types, locations, and adjacency as layout-graph prior | MaskPLAN Our III · RPLAN test | fidimg: 0.241 -> 0.139 | Helped |
| Remove conditioning on previously predicted attributes (ablation) | MaskPLAN · RPLAN test, boundary-only input | fidimg: 4.182 -> 23.103 | Hurt |
| Remove pretrained ADLM (ablation) | MaskPLAN · RPLAN test, boundary-only input | fidimg: 4.182 -> 11.891 | Hurt |
| Replace VQ-VAE with VAE in the ADLM (ablation) | MaskPLAN · RPLAN test, boundary-only input | fidimg: 4.182 -> 39.272 | Hurt |
| Remove per-pixel reconstruction loss (ablation) | MaskPLAN · RPLAN test, boundary-only input | fidimg: 4.182 -> 14.092 | Hurt |
| Use 0%-100% dynamic masking (ablation) | MaskPLAN · RPLAN test, boundary-only input | fidimg: 4.182 -> 6.478 | Hurt |
| Use 25%-100% dynamic masking (ablation) | MaskPLAN · RPLAN test, boundary-only input | fidimg: 4.182 -> 5.169 | Hurt |
| Use 75%-100% dynamic masking (ablation) | MaskPLAN · RPLAN test, boundary-only input | fidimg: 4.182 -> 6.912 | Hurt |
| Increase partial input masking from 20% to 80% | MaskPLAN · RPLAN, partial-input generation | fidimg: 2.314 -> 0.417 | Helped |
| Use complete ground-truth input rather than 80% masked input | MaskPLAN · RPLAN, partial-input generation | fidimg: 0.417 -> 0.593 | Hurt |
Surprising findings
- Feeding 100% of the ground-truth input does not improve over 80% masking (FID 0.417 vs 0.593), which the authors attribute to limited exposure to rich inputs during training.
- Our II with only 25% random partial input surpasses most baselines on FID, and Our III with 60% structured input beats random 60% input on all attributes.
Limitations
- Limited to the RPLAN training dataset, capped at 8 rooms and orthogonal wall arrangements.
- Single-floor layouts only; multi-floor design is not supported.
- Metrics such as MSE over vectorized attributes may not fully capture layout feasibility.
Experimental setup
Datasets: RPLAN
Benchmarks: RPLAN
Metrics: FID on rendered images MSE on room types MSE on adjacency MSE on room areas
Baselines: RPLAN HouseDiffusion iPLAN Graph2Plan
Builds on: Masked Autoencoders (MAE) VQ-VAE Vision Transformer (ViT) Graph-structured MAE
Closest papers by meaning
- OptiScene: LLM-driven Indoor Scene Layout Generation via Scaled Human-aligned Data Synthesis and Multi-Stage Preference Optimization (NeurIPS 2025 · similarity 0.87)
- Visual Layout Composer: Image-Vector Dual Diffusion Model for Design Layout Generation (CVPR 2024 · similarity 0.86)
- Direct Numerical Layout Generation for 3D Indoor Scene Synthesis via Spatial Reasoning (NeurIPS 2025 · similarity 0.86)
- Towards Aligned Layout Generation via Diffusion Model with Aesthetic Constraints (ICLR 2024 · similarity 0.86)
- Constrained Layout Generation with Factor Graphs (CVPR 2024 · similarity 0.86)
- Integrating Symbolic Reasoning into Neural Generative Models for Design Generation (AAAI 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.