- Rating
- 3(+1)
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 3
Read review
Paper summary
The paper proposes ImaGeo, a general-purpose model for single- and multi-view image-to-3D generation. It introduces an attention architecture in which image and geometry tokens are jointly updated and interact with each other, allowing image tokens to evolve during generation and support geometry prediction in the next layer.
Critical strengths
- Simple yet effective. ImaGeo achieves state-of-the-art performance using bidirectional attention and 325K training assets, fewer than several competing methods. The provided qualitative comparisons also show that ImaGeo generates more accurate geometry with richer details than other methods.
Critical weaknesses
- Insufficient experimental evaluation. The paper acknowledges that similar image–shape co-reasoning is used in Hunyuan3D 2.0, so a controlled comparison with its architecture trained on the same data should be included. Moreover, replacing cross-attention with joint attention is expected to substantially increase computational cost, with the gap becoming more pronounced as the number of input images increases. The paper should quantify and report these costs alongside the final generation quality.
Questions for authors
- How much slower is ImaGeo than the cross-attention baseline? Does this gap widen as the number of input images increases?
- Could the authors provide more details on data selection and processing? The paper only states that 325K assets were selected from Objaverse-XL. How exactly were these assets selected and processed? Despite using fewer training assets than TRELLIS and LATTICE, ImaGeo shows clear performance gains, suggesting that data selection and processing may contribute substantially beyond the architectural changes.
3 comments
Thanks for the clarification
Thank you for the detailed response and additional results. My concerns have been addressed, and I have raised my score accordingly.
Thank you for carefully considering our responses and additional results. We are glad that your concerns have been addressed, and we appreciate your constructive feedback, which helped us improve the paper.
Response to Reviewer B28q
Dear Reviewer B28q,
Thank you for your careful review and helpful suggestions. Below, we address your concerns about related architectures, computational cost, and data preparation through the revised experiments and a more detailed description of our data processing pipeline.
Q1. Architecture comparisons
Since related image–shape interaction already exists, how is the architectural contribution established without a Hunyuan-style model trained on the same data?
Response. We add an MMDiT-style + view-ID baseline that removes per-view image self-attention while retaining shared-QKV joint attention (Section 4.4, Table 5). View-ID modulation distinguishes tokens from different views after per-view processing is removed.
With four views and the same Stage 2 model, the Table 3 ImaGeo checkpoint outperforms this Hunyuan/MM-DiT-inspired variant on all five metrics—CD-L1, CD-L2, and normal-map PSNR, SSIM, and LPIPS—on both GSO and Toys4K. For example, PSNR improves by 0.479/0.351 dB, respectively.
Q2. Efficiency and view-count scaling
How much slower is ImaGeo than cross-attention, and how does the cost gap change with the number of input views?
Response. Table 4 shows gains over CA 1.5B on all four quality metrics and comparable quality to CA 5B. Tables 6–7 report Stage 2 DiT costs on one L20 (FP16, 32,768 geometry tokens). With one view, ImaGeo uses only 6.0% more FLOPs and 2.9% more CFG-step time than CA 1.5B, but 46.8% fewer FLOPs and 46.7% less time than CA 5B. Peak allocated memory is 7.756 GiB, versus 8.915/19.563 GiB for CA 1.5B/5B.
From one to six views, ImaGeo’s CFG-step latency increases from 7.855 to 15.969 s. The multi-view cost gap relative to CA remains unmeasured.
Q3. Data selection and preprocessing
How were the 325K assets selected and processed, and could data curation explain the gains beyond the architecture?
Response. We have added the data selection and preprocessing details to Appendix A.4. After initial filtering (~500K assets), we apply Sparc3D-based flood-fill reconstruction [1] and reject meshes with more than 15% internal vertices, retaining ~325K training assets. All architectures in Table 4 use the same training data and preprocessing, controlling for data curation in this comparison.
[1] Zhihao Li, Yufei Wang, Heliang Zheng, Yihao Luo, and Bihan Wen. Sparc3D: Sparse Representation and Construction for High-Resolution 3D Shapes Modeling. NeurIPS, 2025.
Thank you again for your time and consideration. We hope these clarifications and additional results address your concerns. We welcome any further questions and look forward to your feedback.