Native and Compact Structured Latents for 3D Generation
arXiv (Cornell University), 2025 · paper · AI and machine learning
Presents a structured latent representation for 3D generation built on O-Voxel, a sparse voxel format encoding geometry and appearance for arbitrary topology, plus a compressing autoencoder. Flow-matching models with 4B parameters are trained, and the authors say generated geometry and materials far exceed existing models.
Checked: PDF from openaccess.thecvf.com, fetched 5 Oct 2026. Page numbers are those of that PDF.
2 findings
Substance "MD increases by 69% and PSNR decreases 0.5dB at 16× compression, worsening to 526% and 1.6dB at 32×": by Table 3 the 32× MD increase (1.405 → 7.394) is about 426%; 526% is the ratio, not the increase.
- p. 8 Sec. 4.4, Sparse residual autoencoding
worsening to 526% and 1.6dB at 32×
- p. 8 Table 3, rows SC-VAE f32c128 and w/o Residual AE, MD column
7.394
- p. 8 Sec. 4.4, Sparse residual autoencoding
MD increases by 69% and PSNR decreases 0.5dB at 16× compression, worsening to 526% and 1.6dB at 32×
- p. 8 Table 3
SC-VAE f32c128 118 33.9 1.405
Judge's note: Read Table 3 on the page 8 image: f16c32 MD 1.032 vs 1.747 without Residual AE; f32c128 MD 1.405 vs 7.394. At 16× the increase is 69.3%, matching the text's "increases by 69%". By the same definition the 32× increase is 426%; 7.394/1.405 = 5.263 is 526% of the baseline. The sentence therefore mixes two conventions, and rounding cannot bridge the 100-point gap. The PSNR drops (0.53 and 1.64 dB) match the text.
Second check: another model agreed
Report a problem with this finding
Reference or label Introduction gives generation times "∼3s for 512³ … ∼17s for the 1024³ resolution, and ∼60s for the 1536³ resolution on a NVIDIA H100 GPU", while Sec. 4 states "All runtime statistics are reported on an NVIDIA A100 GPU" (Table 1 caption: "runtime measured on an A100 GPU").
- p. 2 Sec. 1, last paragraph
∼60s for the 15363 resolution on a NVIDIA H100
- p. 6 Sec. 4, Implementation details
All runtime statistics are reported on an NVIDIA A100 GPU
- p. 8 Figure 8 labels
20s
- p. 2 Introduction, last paragraph
∼60s for the 1536³ resolution on a NVIDIA H100 GPU
Only the authors can say what is right here.
Judge's note: Read pages 2, 6 and 8. The Introduction names an H100; the implementation details and the Table 1 caption name an A100. The Figure 1 labels (~2s+~1s, ~10s+~7s, ~35s+~25s) sum to the Introduction's 3/17/60 s. H100 is otherwise named only as training hardware in Sec. 4. No statement in the article reconciles the two; the supplement is not among the files. Only the authors can say which GPU the generation timings were measured on. The 20s label in Figure 8 is handled under C013.
Second check: another model agreed
Report a problem with this finding
11 minor (typos and the like; most checked by the judge only)
Wording Figure 8 labels 1024³ generation as "20s" (and "20s+3s" with refinement), while the Introduction gives "∼17s for the 1024³ resolution" (Figure 1: ~10s + ~7s).
- p. 8 Figure 8, 'Scale Up Computation' panel labels
20s
- p. 2 Introduction, last paragraph
∼17s for the 1024³ resolution
Only the authors can say what is right here.
Judge's note: The enlarged crop of Figure 8 shows the labels "1024³ 20s" and "1024³ 20s+3s" in the Scale Up Computation panel. Page 2 gives ∼17s, and the Figure 1 labels sum to 17s. The gap may come from different hardware (H100 in the Introduction versus A100 in the Experiments, see C001) or from different stages being included, but the paper does not say. This is an unexplained discrepancy for the authors to clarify.
Second check: another model agreed
Report a problem with this finding
Wording "our method consistently outperforms all baselines by a substantial margin across every metric" holds only for Ours 1024; on Sketchfab Featured, Ours 512 is below SparseFlex 1024 on All-Surface F1 (0.613 vs 0.684) and Normal PSNR (31.00 vs 32.12).
- p. 6 Sec. 4.1, Shape Reconstruction, second paragraph
our method consistently outper-forms all baselines by a substantial margin across every metric
- p. 8 Table 1, Sketchfab Featured, rows SparseFlex 1024 and Ours 512
0.684
- p. 8 Table 1, Dec. (s) column, rows TRELLIS and Ours 1024
0.108
- p. 8 Table 1, Sketchfab Featured, rows SparseFlex 1024 and Ours 512
0.613
Judge's note: Read Table 1 on page 8. Ours 1024 is best (bold) in every fidelity column. Ours 512 loses to SparseFlex 1024 on the two Sketchfab columns above, where the table underlines the SparseFlex values as second best; it wins the other columns. Some margins are small (Sketchfab outer-surface F1 0.850 vs 0.828). Reading "our method" as the best variant makes the sentence defensible, so this is an overstatement in wording only. The runtime half of the sentence is judged under C009.
Second check: another model agreed
Report a problem with this finding
Wording "despite using only a modest number of tokens and requiring significantly less runtime" is stated against all baselines, but in Table 1 Ours 1024 decodes in 0.301 s versus 0.108 s for TRELLIS.
- p. 6 Sec. 4.1, Shape Reconstruction, 2nd paragraph
despite using only a modest number of tokens and requiring significantly less runtime
- p. 8 Table 1, Dec. (s) column, rows TRELLIS and Ours 1024
0.301
- p. 6 Sec 4.1 Shape Reconstruction
requiring significantly less runtime
- p. 8 Table 1
Ours 1024 9.6K (306K) 16× 0.301
Judge's note: Read the Dec. (s) column of Table 1 on page 8: Dora 37.7/43.0, TRELLIS 0.108, Direct3D-S2 1.86/13.0, SparseFlex 1.22/3.21, Ours 512 0.077 (bold), Ours 1024 0.301. Only Ours 512 is faster than every baseline; Ours 1024, which carries the best fidelity numbers, is about 2.8 times slower than TRELLIS, though much faster than the other baselines. This is a wording overstatement. It is kept separate from C007, which covers the "every metric" half of the same sentence.
Second check: another model agreed
Report a problem with this finding
Wording Table 2 user-study preference shares do not sum to 100%: Pref% 6.40% + 11.8% + 13.3% + 66.5% = 98.0%; Pref-N% 2.82% + 6.57% + 12.2% + 0.469% + 7.51% + 69.0% = 98.6%.
- p. 8 Table 2, Pref% column
13.3%
- p. 8 Table 2, Pref-N% column
12.2%
Only the authors can say what is right here.
Judge's note: Read Table 2 on the page 8 crop. The user-study text on page 7 says only that preference votes were collected from about 40 participants on 100 prompts; it mentions no "none" option or discarded votes. The Pref-N values match vote counts out of 213 (6, 14, 26, 1, 16, 147, totalling 210) and the Pref values match counts out of 203 (13, 24, 27, 135, totalling 199), which suggests 3 and 4 votes fell into an unreported category. This is my inference, not something the paper states. The paper defers details to a supplement that is not among the files, so this is a query to the authors, not a demonstrated error.
Second check: another model agreed
Report a problem with this finding
Typo Reference 28 is cited twice in one group: "[68, 19, 62, 28, 28, 29]".
- p. 3 Sec. 2, Large 3D Asset Generation Models and Systems
[68, 19, 62, 28, 28, 29]
Judge's note: The page 3 image shows "28, 28" in the Sec. 2 citation list, so it is not an extraction artefact. The preceding list in the same paragraph is [68, 58, 7, 19, 62, 28, 29]. Most likely a plain duplicate, though the authors may have intended a different reference.
Second check: none (judge only)
Report a problem with this finding
Typo Figure 3 caption misspells "conversion": "instant bidirectional con-vertion between 3D asset and O-Voxel".
- p. 3 Figure 3 caption
con-vertion between 3D asset and O-Voxel
- p. 3 Figure 3 caption
instant bidirectional con-vertion between 3D asset and O-Voxel
- p. 3 Figure 3 caption
bidirectional con- vertion between 3D asset and O-Voxel
- p. 3 Figure 3 caption
bidirectional convertion
Judge's note: The page 3 image shows "con-/vertion" across the line break in the Figure 3 caption. The body text and Figure 2 use "conversion".
Second check: none (judge only)
Report a problem with this finding
Typo The regularisation term is subscripted with q-hat in Eq. (2) ("λreg d²_q̂") and in the text, but is defined with q-bar: "d²_q̂ = ∥v − q̄∥²".
- p. 3 Eq. (2)
λreg d2q̂
- p. 4 Sec. 3.1.1, text after Eq. (2)
d2q̂ = ∥v − q̄∥2
Judge's note: Page 3 Eq. (2) and the first paragraph of page 4 both show a hat in the subscript and a bar on the right-hand side for the average of the intersecting points. The meaning is clear from the text ("the average of the intersecting points"), so this is a minor notation inconsistency.
Second check: none (judge only)
Report a problem with this finding
Typo "Unlike transformer-based designs in priors work [58, 17]" should read "prior work(s)".
- p. 4 Sec. 3.2.1, first paragraph
Unlike transformer-based designs in priors work [58, 17]
- p. 3 Figure 3 caption
bidirectional con- vertion
- p. 4 Sec 3.2.1
in priors work [58, 17]
Judge's note: Visible on the page 4 image in Sec. 3.2.1. The "convertion" misspelling also mentioned in this cluster is handled under C004.
Second check: none (judge only)
Report a problem with this finding
Typo Mid-sentence capital: "Together, These results demonstrate".
- p. 6 Sec. 4.2, second paragraph
Together, These results demonstrate
Judge's note: Visible on the page 6 image in Sec. 4.2, second paragraph. The preceding sentence also has "ranging from … and …" with no "to".
Second check: none (judge only)
Report a problem with this finding
Typo Baseline name is capitalised inconsistently: "Direct3D-s2 [57]" in the text (Sec. 4.1 and 4.2) versus "Direct3D-S2" in Tables 1 and 2.
- p. 6 Sec. 4.1
Direct3D-s2 [57]
- p. 8 Table 1
Direct3D-S2 512
Judge's note: Page 6 shows "Direct3D-s2" twice in running text; page 8 shows "Direct3D-S2" in both tables. Cosmetic only.
Second check: none (judge only)
Report a problem with this finding
Typo Table 1 caption has subject-verb disagreement: "MD and CD is reported ×10⁶" should be "are reported".
- p. 8 Table 1 caption
MD and CD is reported ×10
- p. 8 Table 1 caption
MD and CD is reported ×106
Judge's note: Visible in the Table 1 caption on the page 8 image. The exponent is 10⁶; the truncated "×10" quote in the cluster is a text-layer artefact. Grammar only.
Second check: none (judge only)