SAM 3D: 3Dfy Anything in Images
Peer-reviewed conference; CVPR 2026 Best Paper Honorable Mention, 2026 · paper · AI and machine learning
Presents SAM 3D, a generative model that reconstructs an object's 3D geometry, texture and layout from a single image, trained with human- and model-in-the-loop annotation. It reports at least a 5:1 win rate over recent work in human preference tests.
Checked: PDF from openaccess.thecvf.com, fetched 5 Oct 2026. Page numbers are those of that PDF.
1 finding
Substance Table 1, Stage 1 Geometry model, SFT row lists datasets 'MITL, SA-3DAO', while the text (Sec. 3.2.2), Fig. 4 ('MITL/ART - 3DO') and Table 4 ('+ SFT (Art-3DO)') name Art-3DO; SA-3DAO is the held-out evaluation benchmark.
- p. 4 Table 1, Stage 1 Geometry model, SFT row
MITL, SA-3DAO
- p. 6 Sec. 3.2.2, Supervised Fine-Tuning
followed by the smaller, high-quality set from 3D artists (Art-3DO)
- p. 8 Table 4
+ SFT (Art-3DO)
Only the authors can say what is right here.
Judge's note: The page 4 image confirms 'MITL, SA-3DAO' in the SFT row. Sec. 3.2.2 (p6) says SFT uses MITL-3DO followed by Art-3DO. Art-3DO is defined in Stage 2.5 as artist-annotated hard cases. SA-3DAO is introduced (p2, p7) as a new 1,000-pair evaluation set. Nothing in the main text says SA-3DAO is used for training. Most likely this is a mislabel of Art-3DO, but only the authors can confirm.
Second check: another model agreed
Report a problem with this finding
9 minor (typos and the like; most checked by the judge only)
Reference or label Sec. 3 says 'Fig. 9b shows that as we run the data engine longer, model performance steadily improves', but Sec. 4.2 attributes that finding to Fig. 9a ('Historical Elo in data engine').
- p. 3 Sec. 3, paragraph after Step 2
Fig. 9b shows that as we run the data engine longer, model performance steadily improves
- p. 8 Sec. 4.2 and Fig. 9 subcaptions
(a) Historical Elo in data engine
- p. 8 Sec. 4.2, first paragraph
near-linear Elo scaling shown in the histori-cal comparisons from Stage 2 of our data engine (Fig. 9a)
- p. 8 Sec. 4.2 and Fig. 9 caption
historical comparisons from Stage 2 of our data engine (Fig. 9a)
Judge's note: p3 cites 9b. On p8, Sec. 4.2 cites Fig. 9a for steady improvement as the data engine runs longer, and Fig. 9b for 'iterating MITL-3DO data alone ... with decreasing marginal impact'. The Fig. 9 caption describes (a) as checkpoints and (b) as SFT with expanded training data. Fig. 9b does show Elo rising with Set 1-4 'additional data from engine', so the pointer is loosely defensible. It is still inconsistent with the paper's own assignment of panels in Sec. 4.2.
Second check: another model agreed
Report a problem with this finding
Wording Fig. 3 caption 'Samples are divided into four rows, based on type.' but the figure groups the four types as a 2x2 grid of quadrants (Isolated, Render-Paste, MITL, Artist).
- p. 4 Figure 3 caption
Samples are divided into four rows, based on type.
Judge's note: In the page 4 image each type label spans a quadrant of two rows by two columns. Every image row mixes two types, so the caption misdescribes the layout.
Second check: another model agreed
Report a problem with this finding
Wording Fig. 7 caption says 'Numbers indicate human preference rates.' but the bars carry no numeric labels; only a 0%-100% axis is shown.
- p. 7 Figure 7 caption
Numbers indicate human preference rates.
Only the authors can say what is right here.
Judge's note: The p7 page image and crop show stacked bars with axis ticks only. No values are printed on or beside the bars. The caption describes content that is absent, and the 5:1/6:1 ratios cannot be read off the figure.
Second check: another model agreed
Report a problem with this finding
Wording Related Work: 'Recent generative methods show strong results on isolated synthetic objects [11, 17]' cites ShapeNet [11] and Objaverse-XL [17], both dataset papers.
- p. 8 Sec. 5 Related Work
Recent generative methods show strong results on isolated synthetic objects [11, 17]
- p. 9 References [11], [17]
Shapenet: An information-rich 3d model repository
Judge's note: In the reference list (p9 and text), [11] is 'Shapenet: An information-rich 3d model repository' and [17] is 'Objaverse-xl: A universe of 10m+ 3d objects'. The citations may be meant to name the synthetic-object datasets. As written, the sentence attributes generative-method results to them.
Second check: another model agreed
Report a problem with this finding
Typo Fig. 2 caption: 'the mixture of transformers architecture apply a two-stream approach' should read 'applies'.
- p. 3 Figure 2 caption
transformers architecture apply a two-stream approach
- p. 3 Figure 2 caption
the mixture of transformers architecture apply a two-stream approach
Judge's note: The page 3 image confirms the subject-verb disagreement.
Second check: none (judge only)
Report a problem with this finding
Typo Sec. 4.1: 'SAM 3D achieves an 5 : 1 head-to-head win rate' should read 'a 5 : 1'.
- p. 7 Sec. 4.1, 3D shape and texture
SAM 3D achieves an
- p. 7
SAM 3D achieves an 5 : 1 head-to-head win rate on real images
Judge's note: Visible in the page 7 image.
Second check: none (judge only)
Report a problem with this finding
Typo Sec. 4.1: 'Fig. 6 and Fig. 21 in the appendix shows qualitative comparisons' should read 'show'.
- p. 7 Sec. 4.1, 3D scene reconstruction
appendix shows qualitative comparisons
Judge's note: Confirmed in the page 7 image: the compound subject takes a singular verb.
Second check: none (judge only)
Report a problem with this finding
Typo End of Sec. 4.1: 'annotators preferred outputs using an improved estimator not seen during training' has no terminal full stop before the '4.2. Analysis Studies' heading.
- p. 8 End of Sec. 4.1, before '4.2. Analysis Studies'
outputs using an improved estimator not seen during training
- p. 8 Sec. 4.1, end of 3D scene reconstruction paragraph
annotators preferred outputs using an improved estimator not seen during training
Judge's note: The page 8 image shows no period after 'training'.
Second check: none (judge only)
Report a problem with this finding
Typo Reference [122] (Hi3dgen): 'arXiv preprint arXiv:2503.22236, 3:2, 2025.' contains a stray volume/page '3:2'.
- p. 13 Reference [122]
arXiv preprint arXiv:2503.22236, 3:2,
Judge's note: Visible in the p13 crop. An arXiv preprint entry has no volume or page, so '3:2' is a bibliography artefact.
Second check: none (judge only)