GREPGrounding Generated Room
Evaluation in Principles

Mohan Cao, Yongsen Mao, Chucheng Xiang, Zhongyuan Liu, Zhihang Zhong, Haiping Wang, Yuan LiuCorresponding authors

GREP evaluates generated 3D rooms through Physical Validity, Visual Coherence, and Human-Use—and grounds its judgments in scene evidence.

87% agreement with human preferences across 15 scene pairs vs 62% for a holistic VLM ↗
A sofa resting on a rug, with the contact circledA Physical ValidityIs the contact intended? A pair of oversized shoes circled on the floor of a generated roomB Visual CoherenceIs the scale plausible? Two pairs of chairs facing each other with almost no gapC Human-UseCan people use it?

1. What GREP does

Given a furnished 3D scene, GREP evaluates its quality and explains the problems it finds.

Diverse Scene Generators

Models · Pipelines · Agents

A furnished 3D scene submitted for evaluationFurnished 3D scene

GREP evaluation

3 principles · 8 metrics

A Physical Validity

  • Collision
  • Support
  • Out-of-bounds

B Visual Coherence

  • Scale
  • Style
  • Pairing

C Human-Use

  • Placement
  • Function
ScoresOverall · Principles · Metrics
Diagnostic reportAffected objects · Reasons · Evidence

2. Why do we need GREP?

Rooms need contextual judgments—and the right questions.

Paper Figure 1: sofa–rug compression and table–rug penetration, each shown in a zoomed inset
Context matters

Contact is not always a defect

SceneEval
Flags both rug contacts.
GREP
Clears sofa compression; confirms table penetration.
Paper Figure 1: the scene view, and the judge’s view showing facing chairs 0.16 metres apart
Questions matter

A plausible room can be unusable

Other evaluators
The global VLM says “looks fine”; SceneEval and LEGO-Eval miss this pair.
GREP
The 0.16 m gap leaves no room to sit.
Original examples and comparisons · Figure 1 ↗

3. How does GREP do it?

Principled, adaptive discovery

The room determines what to check

Principles and check types stay fixed. GREP selects the objects and relationships that warrant inspection—including the facing chairs above.

Check instantiation · Appendix A ↗
Proactive evidence

Insufficient evidence prompts another look

A 173° angle cannot identify the desk’s usable side. GREP requests two views and finds the chair at its closed back.

Evidence operations · Paper p. 4 ↗
A measured angle of 173 degrees is inconclusive. Two new views reveal blocked knee space and a chair at the desk's closed back.
Measurement → insufficient evidence → new views → judgment. Figure 1 ↗

4. From principles to judgments

The principles guide what to check. The evidence determines the judgment.

  1. 01

    Formulate checks

    Guided by the principlesA Physical ValidityB Visual CoherenceC Human-Use

    Identify the objects and relationships that need inspection in this room.

    Can the cabinet door open fully?
  2. 02

    Inspect & judge

    Measure geometryRetrieve existing viewsRender new views
    Evidence sufficient?
    No Acquire more evidence
    and inspect again
    Yes →Make a judgment

    Budget reached? Record the unresolved gap.

  3. 03

    Score & diagnose

    Scores

    Overall · Principles · Metrics

    Diagnosis

    Affected objects · Reasons · Evidence

Follow a check in the room ↓Original framework · Figure 2 ↗

5. Inspect the room

Start with the cabinet. Follow the evidence, then compare a revised placement.

Illustrative scene, not a benchmark output

Choose a check

Human-Use
1 Formulate checks

Can the cabinet door open fully?

  1. 1Formulate checks
  2. 2Inspect & judge
  3. 3Score & diagnose
Room overview · Drag to explore
Static preview of the illustrative living room with a sofa, desk and cabinet
Preparing the room

The interactive view will appear when loading finishes.

GREP camera
Original arrangementDrag to orbit · Scroll to zoom
Scene asset credits ↗

Explore evaluated scenes

6. What GREP reveals

Agreement with people—and differences in how generators use guidance.

86.7%GREP
62.2%Holistic VLM
53.3%Physical-only

15 pairs · 6 scenes · 5 raters. Holistic VLM averaged over three runs. Figure 6 ↗

FrozenAssets* excludes Visual Coherence from Overall. Integration is a research direction, not a capability demonstrated here.

7. Bonus: reports for revision

A diagnosis gives the generator a concrete problem to address.

0.585 → 0.676

Mean score after three report-guided revisions. Score-only and no-feedback branches decline.

Five scenes · GPT-4.1 · Separate 0–1 refinement protocol; these are not leaderboard scores divided by 100.

Reports changed the layout in 15/15 revisions, versus 5/15 with scores alone.

Original Figure 5a: report-guided revision improves the mean from 0.585 to 0.676, while scores-only and no-feedback decline.
Original revision curve · Figure 5a ↗
A real diagnostic record · S104 revision study

From the paper’s Table 9; separate from the selected benchmark example.

Metric / check
Function · Relative-use geometry
Objects
Piano bench · Digital piano
Severity
Blocked
Reason
The bench is too far from the piano to serve as usable seating.
Measurement
Centre distance: 2.0 m. The bench’s seating side and the piano’s control side are each at 90° to the other object’s direction.
Table 9 · Full source ↗

Citation

If you find GREP useful, please cite arXiv:2611.05802.

@article{cao2026grep,
  title   = {GREP: Grounding Generated Room Evaluation in Principles},
  author  = {Cao, Mohan and Mao, Yongsen and Xiang, Chucheng and Liu, Zhongyuan and
             Zhong, Zhihang and Wang, Haiping and Liu, Yuan},
  journal = {arXiv preprint arXiv:2611.05802},
  year    = {2026}
}