Roomer: Reflective Object-Grounded Model Editing and Repair for 3D Indoor Layout Synthesis
Lingwei Dang, Ziyan Qiu, Jiajia Cheng, Shishuo Shang, Zhenhao Zhang, Yufei Zhu, Qingxin Xiao, Pan Liu, Shenghui Huang, Yun Hao, Juntong Li, Qingyao Wu
cs.RO, cs.CV
2026-08-03
Generated room layouts keep local violations; Roomer locates the offending furniture like a transaction, applies a minimal verified patch, and cuts ReSpace out-of-bounds from 14.7% to 1.2%.
Today's indoor layout generators produce rooms that look broadly plausible but hide local violations on close inspection: furniture interpenetrating and colliding, pieces placed out of bounds, openings or circulation blocked. Most prior work does full-scene synthesis or scene-level optimization, and rarely pinpoints which specific object is at fault so that only the affected region gets repaired. Regenerating the whole scene risks disrupting already-valid furniture and introducing new violations. Roomer aims at local, object-level repair: find the offending piece, make a minimal change, and verify the fix actually worked without breaking anything else.
Roomer is a reflective repair framework that borrows the commit-or-rollback logic of a database transaction.
On Common-1100 (density-controlled comparison):
| Method | OOB down | COL down | Practical up |
| InstructScene | 37.49% | 28.50% | 40.99% |
| ReSpace | 14.72% | 36.56% | 66.43% |
| Roomer-initial | 24.40% | 29.52% | 72.50% |
| Roomer-final | 8.66% | 17.50% | 82.98% |
The plug-and-play result is more telling. Attached as a post-processing layer to other generators, Roomer cuts ReSpace's out-of-bounds rate from 14.72% to 1.16%, collisions from 36.56% to 7.61%, and lifts Practical from 66.43% to 75.12%; DiffuScene, InstructScene, and SemLayoutDiff show similar gains. In the ablation, removing RoReview drops the Hard repair rate to 19.73%; the full model reaches 72.64% Hard, 82.98% Practical, and 96.27% Safe. A professional blind review approves the high-Practical tier 83.3% of the time (Z=4.67, p=3.07e-6).
Roomer is pragmatic about its scope. It does not compete on who generates the best layout; it acts as a universal post-processing repair layer for any existing generator, and it touches only the broken region, leaving valid furniture alone. The transactional commit-or-rollback imports a proven idea from software engineering to guarantee monotone improvement with no new violations. For 3D scene generation work, this means a local bug need not trigger retraining the whole generator; attaching Roomer is enough.
The authors concede that Roomer covers only a predefined residential rule set and searches within finite, action-specific candidate spaces, so issues outside that coverage or without an acceptable candidate go unrepaired; they plan to extend the rule library and candidate generation. One open question: Practical is a composite metric defined in the paper, and exactly how the 82.98% is computed and how it maps to real usability sits in the appendix. Repair is also iterative over multiple rounds (Figure 4 shows adaptive calls concentrating on a shrinking set of unresolved scenes), and the paper does not give clear numbers on latency or the round bound for very complex layouts.