- Rating
- 2
- Confidence
- 3
- Informativeness
- 1
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This work proposes a method for floorplan generation conditioned on room adjacency graphs and outer boundaries. It additionally proposes conditioning mechanisms for room sizes and window/door positions. The title and abstract promise a “constraint-consistent diffusion framework” which could be understood as strictly enforced constraints, but they are soft constraints implemented through guidance. Further, the method proposes a simple workaround for a limitation inherited from HouseDiffusion that requires knowing the number of room corners in advance of the floorplan generation.
Critical strengths
None
Critical weaknesses
A key claimed contribution of the paper is the “polygon augmentation strategy”. The baseline HouseDiffusion (CVPR 2023) assumes that the number of per-room corners is known in advance of the diffusion process (from dataset room-type statistics). This is a limitation that the proposed method tries to solve via data augmentation. I.e. it adds redundant points that can then simply be placed on a straight line if not needed. While effective, it does not seem particularly elegant (this is my subjective opinion, of course). Other methods actually learn the number of room corners (see RoomFormer uses a parallel Transformer, SceneScript an auto-regressive Transformer) while this method learns to become more robust to a wrong given number of corners. A more elegant approach would be to learn which corners are indeed needed via an additional classification score. This could be a meaningful baseline, and also result in a more compact vectorized representation getting rid of redundant points.
The paper’s other claimed contribution is the joint conditioning on a number of properties, including the room size and door/window positions. While this work is the first to propose that specific task, the technical implementation is not evaluated clearly. For example, the proposed spatial conditioning represents boundaries by adding 8 additional vertices along each edge, which are then encoded with a “Spatial Encoder” attended to via cross-attention. GSDiff proposes an alternative approach that renders the boundary to a 256x256 image that is then attended to via cross-attention. This work only explains what they do, but there is no comparison to prior approaches -- it would be good to have an analysis that compares different ways of providing the conditions to know which one to choose.
Questions for authors
How would this method compare to a baseline that learns the number of corners instead of doing data-augmentation to become robust towards a wrong initial corner-number?
The work proposes different conditionings, but the actual conditioning mechanisms are not compared to prior work - e.g. the graph and the boundary conditioning. Can such a comparison be provided?
Remarks: l.61 - what are “vector diffusions”? Isn’t every diffusion on a vector? Fig. 2: The comparison of different methods with different conditionings seems somewhat misleading. They should be split into different Figures depending on the conditioning types.
2 comments
Q2. How does the conditioning mechanism compare with prior work?
We did not ablate the graph-conditioning mechanism because graph-guided attention is central to the original HouseDiffusion architecture, which we use as our backbone. Replacing it would change how the backbone models room relationships. We therefore retain this mechanism as designed and focus our ablation on the spatial conditioning introduced in this work.
We compared three ways to give the denoiser the apartment geometry. In our original model, a spatial encoder represents each boundary segment as a token. Every floorplan vertex can attend to these tokens through a dedicated spatial room cross-attention layer (SRCA). Our first alternative uses the same encoded segments but appends the boundary tokens to the sequence of floorplan vertices. Global self-attention then lets the vertices and the boundary tokens attend to one another. This changes how the boundary enters the denoiser while keeping the SRCA pathway for windows and the entrance door.
Our second alternative takes the spatial condition as an image. Raster boundary encoding has precedent in Graph2Plan, which uses a CNN. For our diffusion model, we render the boundary and any supplied openings in separate channels of a single 224 × 224 image. A frozen DINOv2 ViT-B/14 encodes the image into 256 patch tokens, and every floorplan vertex can attend to every patch through cross-attention.
We evaluated checkpoints on the same setup as in Table 1 of the main paper. CFG 0 disables target room areas and CFG 3 enables size guidance.
| Size CFG | Spatial conditioning | FID ↓ | FIDgs ↓ | KID ↓ | KIDgs ↓ | Boundary IoU ↑ | GED ↓ | Edge IoU ↑ | SizeError ↓ |
|---|---|---|---|---|---|---|---|---|---|
| 0 | Vector segments + SRCA (ours) | 6.43 | 4.21 | 0.89 | 0.63 | 0.966 | 0.330 | 0.930 | 0.163 |
| 0 | Vector segments + boundary concatenation | 6.52 | 4.33 | 1.13 | 1.02 | 0.968 | 0.296 | 0.938 | 0.160 |
| 0 | Raster + DINOv2 cross-attention | 6.21 | 4.20 | 0.64 | 0.63 | 0.957 | 0.287 | 0.931 | 0.163 |
| 3 | Vector segments + SRCA (ours) | 6.05 | 3.98 | 0.96 | 0.52 | 0.974 | 0.282 | 0.942 | 0.028 |
| 3 | Vector segments + boundary concatenation | 5.99 | 4.15 | 0.70 | 0.72 | 0.976 | 0.258 | 0.946 | 0.021 |
| 3 | Raster + DINOv2 cross-attention | 6.07 | 3.87 | 0.74 | 0.77 | 0.967 | 0.271 | 0.942 | 0.023 |
The three conditioning mechanisms achieve broadly comparable results on this benchmark. At CFG 3, FID ranges from 5.99 to 6.07 and Boundary IoU from 0.967 to 0.976. Boundary concatenation gives the highest Boundary IoU at both guidance scales, while raster conditioning gives the lowest KID and GED at CFG 0. These results support all three designs as viable ways to condition a vector floorplan generator on spatial constraints.
Q3. Vector diffusions
By “vector diffusions,” we meant diffusion models that directly denoise polygon vertex coordinates. We will use this more precise wording in the final version.
Q1. How would this method compare to a baseline that learns the number of corners instead of doing data-augmentation to become robust towards a wrong initial corner-number?
We trained a separate predictor to estimate each room’s corner count before generating the floorplan. We initialized it from the coordinate diffusion model trained without corner augmentation and fine-tuned it for 10 epochs. The predictor uses one token per room, encoding its type, identity and target area. Each room attends to all rooms in the apartment, to its neighbors in the adjacency graph, and to the exterior boundary through cross-attention. A linear head predicts one scalar per room. We trained it with mean squared error against the ground truth corner count, then rounded each prediction to the nearest integer with a minimum of four. We supplied these budgets to the diffusion generator trained without corner augmentation.
Corner-count prediction. On the 600-plan RPLAN split used for predictor validation (4,816 rooms), the predictor achieved a mean absolute error of 0.575 corners per room before rounding and 0.506 after rounding. Always allocating four corners gave an MAE of 1.216.
Floorplan generation. We compare methods on the same data as in Table 8 of the appendix. All rows use room-size conditioning with guidance scale 3.
Even at half a corner of error per room, replacing the true counts with learned counts increased FID from 4.71 to 14.73 and size error from 0.017 to 0.030. Table 8 of the appendix shows how augmentation changes the generator’s response to surplus vertices. From Ground truth vertices (Original) to Extra vertices (More Points), FID changes from 5.38 to 5.41 with augmentation and from 4.71 to 16.23 without it.
When ground-truth corner counts are unavailable at inference, the augmented generator with a fixed budget (N=8) outperforms the generator using learned counts on all six metrics in this experiment. For example, it achieves a FID of 5.45 versus 14.73, a GED of 0.22 versus 0.34, and a size error of 0.018 versus 0.030. The fixed budget also removes the need to train a separate corner-count predictor.