- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
ScenePilot uses a shared Qwen3-VL-4B policy to repair indoor scenes through a canonical object-centric representation and backend adapters. It proposes atomic edits, retaining them only when a verifier’s combined physical, spatial and visual-semantic score improves. The policy is trained with SFT and GRPO on SceneReverse recovery trajectories constructed from perturbed 3D-FRONT rooms. The authors report 4–33% Overall score gains across five generators on 100 SceneEval prompts using one frozen policy.
Critical strengths
- The problem framing and interface design are clear. The canonical state, capability descriptor and accept-or-rollback verification are sensible abstractions.
- SceneReverse is documented with room-disjoint splits and action statistics.
- The evaluation uses paired comparisons across five generators and reports individual SceneEval metrics, including decreases in support and accessibility on ReSpace and Reason-3D.
Critical weaknesses
-
Eq. 7 leaves Q_PHY, Q_SPA, normalization, ε and inference weights unspecified. Table 2 shows that training and verification help, but lacks a random-edit or collision-resolution proposer under the same verifier and budget. This control is particularly important if verification reuses SceneEval’s evaluation components.
-
No repair comparisons are made against SceneReVis [1], LayoutVLM’s optimization stage, DiffuScene [2], PhyScene [3] or DeBaRA [4] on shared starting scenes. GPT-5.4 outperforms ScenePilot by 6.54 Overall points, but no quality-cost comparison is provided. The inference judge, call counts and per-scene time/cost are unspecified. Using GPT-5.4 for both reward and evaluation also raises possible judge-specific bias despite the Gemini check.
-
T_max, repair rounds, acceptance/rollback rates and edit magnitudes are unreported. If Table 10 covers complete runs, only 2 to 12 of 100 terminate through STOP, suggesting budget-driven termination. Support and accessibility decline on ReSpace and Reason-3D, so Overall gains do not establish conservative repair.
-
The policy’s prompt and history inputs differ across Fig. 2, Eq. 3, Algorithm 1 and G.2; instruction pairing in SceneReverse is also unexplained. Sec. 4.1 names five unseen generators, while A.4 lists four and clarifies tuning exclusion only for LayoutGPT. Policy/verifier inputs, each generator’s tuning status and the 100 evaluation prompts need clarification.
-
The perceptual results are inconsistent across tables, and the human–VLM agreement statistics do not match the underlying ratings. The evaluation cohorts and denominators need clarification. Agreement is measured only across condition averages, while the human study’s scene count and statistical uncertainty are unreported.
Minor
- Table 4 calls populated hyperparameters “placeholders” despite describing a final run.
- The GRPO judge prompt mixes “Be generous” and “7 to 10” with a full 0 to 10 rubric; reward distributions and within-group variance would clarify the resulting signal.
References
[1] Zhao, Yang, et al. "SceneReVis: A Self-Reflective Vision-Grounded Framework for 3D Indoor Scene Synthesis via Multi-turn RL." arXiv preprint arXiv:2602.09432 (2026).
[2] Tang, Jiapeng, et al. "Diffuscene: Denoising diffusion models for generative indoor scene synthesis." 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2024.
[3] Yang, Yandan, et al. "Physcene: Physically interactable 3d scene synthesis for embodied ai." 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2024.
[4] Maillard, Léopold, et al. "Debara: Denoising-based 3d room arrangement generation." Advances in Neural Information Processing Systems 37 (2024): 109202-109232.
Questions for authors
Refer to weaknesses.