An AI system for generating roof structures beat one competing graph generator on all five reported unconditional metrics, while a different system was strongest overall when roofs were reconstructed from images. RoofDiT also operates on planar roof graphs and does not directly recover full three-dimensional geometry, roof heights or watertight building models.
How the model draws a roof
RoofDiT works with planar roof graphs, layouts represented with nodes and edges. Its two-stage pipeline first uses a conditional diffusion transformer to generate the nodes, then an edge-prediction module chooses connections from that generated set. Footprint information and aerial images can be supplied as optional conditions.
The node-generation objective combines diffusion loss with a roof-alignment term. The data came from the residential roof dataset introduced by Ren et al., with annotated roofs and corresponding aerial images. The reported split included 1,926 training samples, 249 validation samples and 223 test samples.
A clear win in free generation
For the unconditional test, the study generated 1,000 roof graphs and compared them with repeated balanced subsets of the test set using FID and KID. RoofDiT recorded a roof-valid rate of 0.943 against 0.941 for GSDiff, and a planar rate of 0.958 against 0.952. Its duplicate rate at the reported @3 setting was 0.002 against 0.004, while FID was 45.329 against 53.954 and KID was 23.442 against 31.210.
Taken together, those five reported measures favored RoofDiT: its validity and planarity figures were higher, while its duplicate, FID and KID figures were lower. But the analysis is descriptive rather than a statistical verdict, reporting no confidence intervals, repeated-run variability or significance tests. The result shows a lead in this evaluation without quantifying how stable the gap would be across repeated experiments.
The footprint changes the tradeoff
With a roof footprint supplied, RoofDiT produced five outputs per test footprint, allowing the study to report both averaged results and a best-of-five result. Its edge F1 was 0.981, compared with 0.960 for the straight-skeleton baseline. Within RoofDiT's own results, best-of-five face F1 was 0.886 versus 0.798 for the ordinary figure, and best-of-five valid rate was 0.930 versus 0.841.
That advantage came with a tradeoff. The straight-skeleton baseline remained stronger on node matching F1, face F1, matched IoU, planar rate and valid rate, even though RoofDiT had the higher edge F1. A best-of-five score should be read as evidence that one good candidate was present among five, not as proof that a single output will reliably match the target.
Images produce a more mixed result
Within RoofDiT's image-guided tests, the version given both an aerial image and a footprint scored higher on the reported matching measures than the aerial-only version. Node F1 was 42.6 with aerial imagery alone and 70.3 with imagery plus footprint; edge F1 was 78.4 versus 98.4, and face F1 was 63.5 versus 82.4. Node-count MAE was 2.23 versus 0.90, while valid rate was 87.4 versus 88.3 and planar rate was 90.1 versus 90.6.
Against the selected image-guided baselines, however, HEAT had the strongest overall reconstruction performance, while RoofMapNet performed substantially worse across metric groups. RoofDiT was competitive and particularly strong on edge recovery, especially with footprint input, but the comparison is not a clean head-to-head: HEAT and RoofMapNet use pixel-aligned roof annotations, whereas RoofDiT uses canonicalized graph representations. HEAT was retrained on the same split; RoofMapNet used a released pretrained model.
A planar map, not a 3D building
These caveats narrow the claim. RoofDiT is a competitive option for planar roof-graph generation and footprint-augmented reconstruction, but the results do not establish overall superiority for image-guided reconstruction. The model does not directly recover full three-dimensional geometry, roof heights or watertight building models, and its validity is learned rather than guaranteed by construction.
Paper data and sources
Original title: Diffusion Transformers for Roof Graph Synthesis and Reconstruction
Authors: Daniel Panangian, Ksenia Bittner
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text