- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper focuses on recovering a complete scene mesh from a partial point cloud. Built on the TRELLIS architecture, the main technical contributions are: (1) partial-point conditioning — a frozen NOVA3R encoder produces global tokens injected via cross-attention, plus a local path that encodes the voxelized input into the SS latent, replaces the initial noise at visible cells, and concatenates the visibility mask at every step; (2) a Swin-DiT backbone (3D window attention implemented with masks), which makes a 32³ latent / 128³ occupancy grid fit in memory during training. The authors train on InternScenes and ScanNet++ and evaluate on Replica and SCRREAM. The results show a large lead over SG-NN, NKSR and Seen2Scene. The appendix adds a symmetric Chamfer, results for one checkpoint across seven input sources, and failure cases.
Critical strengths
1.Experimental results are much better than the baselines, and the design of how to inject partial scene point cloud, as well as the swin-dit are reasonable and align with the target.
- The ablation earns its place. Removing local injection is what breaks the model (CD_all 0.071 → 0.105), while dropping the global tokens costs much less (0.086), so per-voxel anchoring is clearly the load-bearing idea. Tab. 2b is candid about the Swin-DiT price, including the parts that don't flatter it: at 16³ full attention wins CD_all and [email protected], and Swin only pulls ahead at 2 cm.
Critical weaknesses
1.The comparison and discussion of closely related generative reconstruction methods are limited, particularly ReconViaGen and GenRecon. ReconViaGen also conditions TRELLIS generation on reconstruction information, while GenRecon extends an object-level generative prior to scene reconstruction. GenRecon is cited, but the paper offers neither a detailed methodological comparison nor an experimental one. These methods primarily use image conditioning rather than point-cloud input, and ReconViaGen targets objects, so they are not directly interchangeable baselines. I would like the authors to clarify the differences in conditioning, observation consistency and scene scaling, and add experiments where comparable inputs and implementations permit. This would make the contribution relative to this line of work clearer.
2.Table 2b supports the efficiency advantage of Swin-DiT, but the source of its quality gains could be better understood. Table 2a compares full attention at 16³ with Swin-DiT at 32³, changing both architecture and resolution. Reporting quality for Swin-DiT at 16³ would allow a same-resolution comparison with full attention. This would strengthen the ablation analysis without changing the existing evidence that Swin-DiT reduces computational cost and enables higher-resolution training.
Questions for authors
-
What are the main differences from ReconViaGen and GenRecon in conditioning, observation consistency and scene scaling? Which follow from the input setting, and which are methodological?
-
Could you add a scene-level comparison with GenRecon using the same input views and evaluation protocol, while specifying the conditioning information each method uses? If this is currently infeasible, what prevents it?
-
How does Swin-DiT at 16³ compare in quality with full attention at the same resolution?