A Style-Based Generator Architecture for Generative Adversarial Networks
2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019 · paper · AI and machine learning
Proposes a redesigned generator for generative adversarial networks, inspired by style transfer, that separates high-level attributes from stochastic variation and allows scale-specific control. It reports better quality metrics, introduces two new measures and a new face dataset.
Checked: PDF from arxiv.org, fetched 4 Oct 2026. Page numbers are those of that PDF.
0 findings
None that passed a second check.
7 minor (typos and the like; most checked by the judge only)
Wording Sec. 2.1 says the style-based generator (E) improves FID over B by "almost 20%", but Table 1 gives 6.11→5.06 (CelebA-HQ) and 5.25→4.42 (FFHQ), which are about 17% and 16%.
- p. 3 Section 2.1, left column
almost 20%, corroborating the large-scale ImageNet mea-
- p. 2 Table 1, rows B and E
5.06
- p. 2 Table 1, rows B and E
6.11
Judge's note: Table 1 shows B = 6.11/5.25 and E = 5.06/4.42. The relative improvements are 17.2% and 15.8%. No pairing of B with E or F reaches about 20%. "Almost 20%" is a loose round-up, so this is a wording point and the data are fine.
Second check: another model agreed
Report a problem with this finding
Typo Figure 1 caption: "feeds the latent code though the input layer only" — "though" should be "through".
- p. 2 Figure 1 caption
feeds the latent code though the input layer only
- p. 2 Figure 1 caption, line 1-2
though the input layer only
Judge's note: Page image crop shows "though the input layer only" in the Figure 1 caption. This is a real typo, not an extraction artefact.
Second check: none (judge only)
Report a problem with this finding
Typo Sec. 2.2: "while the latent is still provided though the input layer" — "though" should be "through".
- p. 3 Section 2.2 Prior art, second paragraph
while the latent is still pro-vided though the input layer
- p. 3 Section 2.2, second paragraph
vided though the input layer
- p. 3 Section 2.2
provided though the input layer
Judge's note: The page image confirms "pro-vided though the input layer" in Section 2.2.
Second check: none (judge only)
Report a problem with this finding
Typo Appendix E: "Appendix Bwith ψ = 0.7" — a space is missing between "B" and "with".
- p. 9 Appendix E, first paragraph
Appendix Bwith ψ = 0.7
- p. 9 Appendix E, first paragraph
Appendix Bwith
Judge's note: The page image shows "Bwith" run together, so this is in the PDF and not lost spacing in the text layer.
Second check: none (judge only)
Report a problem with this finding
Typo Appendix E names the datasets "LSUN BEDROOM, CARS, and CATS", while the Figure 11/12 captions say "LSUN CAR" and "LSUN CAT".
- p. 9 Appendix E
BEDROOM, CARS, and CATS, respec-tively
- p. 10 Figure 11 caption
with the LSUN CAR dataset at 512 × 384
- p. 11 Figure 12 caption
with the LSUN CAT dataset at 2562
Judge's note: Confirmed in the images: the text uses CARS/CATS and the captions use CAR/CAT. This is a cosmetic inconsistency, and the meaning is clear.
Second check: none (judge only)
Report a problem with this finding
Typo Reference [23] author field reads "W.-S. Z. Guang-Yuan Hao, Hong-Xing Yu.", which is malformed and inconsistent with the other entries.
- p. 11 Reference [23]
W.-S. Z. Guang-Yuan Hao, Hong-Xing Yu. MIXGAN
- p. 11 Reference [23]
W.-S. Z. Guang-Yuan Hao, Hong-Xing Yu.
Only the authors can say what is right here.
Judge's note: Rests on outside knowledge: the correct author list is G.-Y. Hao, H.-X. Yu and W.-S. Zheng. The page image confirms the garbled author field, which puts a truncated "W.-S. Z." first and uses full given names, unlike the "Initials Surname" format of every other entry.
Second check: none (judge only)
Report a problem with this finding
Typo Reference [14] gives the author as "H. Drucker and Y. L. Cun.", splitting the surname LeCun.
- p. 11 Reference [14]
H. Drucker and Y. L. Cun.
Only the authors can say what is right here.
Judge's note: Rests on outside knowledge: the second author of the double backpropagation paper is Yann LeCun. The page image shows "Y. L. Cun".
Second check: none (judge only)