You can deploy this model and start earning money today!

gafua926/
outlineflow

Name: outlineflow
Availability: InStock

Views

No views yet

OutlineFlow — outline → typed room polygons

A from-scratch rectified flow-matching set-Transformer that, given only an apartment/floor outline, generates a complete set of typed interior room polygons (vector, not pixels). Challenge: Mirror Mirror on the Wall (DAVIS × Paris 2026), dataset: Modified Swiss Dwellings.

Final results (full plan_id model, align g16 + churn 0.3, 3-seed mean)

Config	FID ↓	Density ↑	Coverage ↑
Raw baseline	167	0.097	0.067
Final	135.3	0.088	0.111

FID −19%, Coverage +66% over baseline.

Usage

python

from generate import generate          # from the code repo
rooms = generate(outline)              # outline: shapely (Multi)Polygon or WKT
# rooms == [(shapely Polygon, room_type_id), ...]   (vector, gap-free, seed 42)

Load weights: outputs_full_plan_id/ckpt.pt (the default checkpoint generate() reads).

Method (short)

Pure rectified flow matching; DiT-style set-Transformer (AdaLN-Zero), outline encoded by a permutation-invariant PointNet, no positional encoding (a plan is a set).
Diagnosis: the L2 objective regresses to the conditional mean → variance collapse → weak Density/Coverage.
Two levers that work: grid-align post-process (FID) + churn / SDE sampling (Coverage).
Diffusion (DDPM) and EBM variants were tested and did not beat this; richer conditioning (cross-attention + scale) raised outline-reading r 0.06→0.67 but did not move the metric, confirming the bottleneck is the loss, not the conditioning.

Code & full write-up: see the GitHub repo (training/eval/generation + result.md).

d_model=128, L=4, 1.33M params, 15k steps, full data (5,372 plans / 203k rooms), seed 42.

gafua926/outlineflow

OutlineFlow — outline → typed room polygons

Final results (full plan_id model, align g16 + churn 0.3, 3-seed mean)

Usage

Method (short)

gafua926/
outlineflow