High-Resolution Image Synthesis with Latent Diffusion Models

• p. 6 Table 2, row label and caption

  †/∗:Numbers from [109]/ [26]

Only the authors can say what is right here.

Judge's note: Read in the p.6 image. ∗ also marks GLIDE∗ and Make-A-Scene∗. The caption gives ∗ no other meaning, so the marker on the authors' own row is unexplained. It is probably meant to indicate guidance.

Second check: another model agreed

• p. 28 Table 18, LSUN Churches row

  LDM-8 (ours, 100 steps, 410K)

• p. 6 Table 1, LSUN-Churches

  LDM-8∗(ours, 200-s)

• p. 41 Figure 30 caption

  η = 0 (FID = 4.48)

• p. 6 Table 1

  LDM-8∗(ours, 200-s) 4.02

Only the authors can say what is right here.

Judge's note: Page 6 crop: Table 1 Churches row reads LDM-8* (ours, 200-s) 4.02 / 0.64 / 0.52. Page 28: Table 18 gives the same metrics with '100 steps'. The two step counts for one result need clarifying. On Fig. 30 (page 41): the η = 0 captions differ from Table 1 for both CelebA-HQ (Fig. 28: 5.15 vs 5.11) and Churches (4.48 vs 4.02), while the η = 1 captions (Figs. 29, 31) match. The caption FIDs may therefore come from different sampling settings, and that part is a question for the authors rather than a clear error.

Second check: another model agreed

• p. 7 Sec. 4.3.1, last paragraph

  els with f ∈{4, 8} from Sec. 4.1 in Tab. 3

• p. 16 Appendix A changelog

  class-conditional synthesis on ImageNet in Sec. 4.1, Tab. 3

• p. 16 Appendix A Changelog, 2nd bullet

  We updated results on class-conditional synthesis on ImageNet in Sec. 4.1, Tab. 3

• p. 7 Sec. 4.3.1, last paragraph

  models with f ∈ {4, 8} from Sec. 4.1 in Tab. 3, Fig. 4 and Sec. D.4

Judge's note: Tab. 3 is on p.7 and is discussed in the last paragraph of Sec. 4.3.1. Sec. 4.1 covers the perceptual-compression trade-offs (Figs. 6/7). The Sec. 4.3.1 wording "models with f ∈ {4, 8} from Sec. 4.1 in Tab. 3, Fig. 4 and Sec. D.4" is defensible: it names the f values chosen in Sec. 4.1, and LDM-8 appears in D.4/Tab. 10. Only the changelog section pointer is off.

Second check: another model agreed

• p. 28 Table 18, ImageNet

  LDM-8-G (ours, 100, 2.9M)

• p. 22 Table 10

  200 DDIM steps, classifier scale 10, 2.9M train steps, batch size 64

• p. 22 Table 10, LDM-8-G (2.9M) row

  200 DDIM steps, classifier scale 10, 2.9M train steps

Only the authors can say what is right here.

Judge's note: Page 22 image: Table 10 row LDM-8-G reads 8.11, 190.43, 0.83, 0.36, 506M, '200 DDIM steps'. Page 28: Table 18 shows the same metrics with '100'. The throughput could have been measured at a different step count, but the row label does not say so. The authors need to clarify.

Second check: another model agreed

• p. 6 Text below Table 2

  LDMs consistently improve upon GAN-based methods in Precision and Recall

• p. 6 Table 1, LSUN-Bedrooms

  StyleGAN [41] 2.35 0.59 0.48

Judge's note: Read in the Table 1 image (p.6). LDM precision and recall exceed the GAN values on FFHQ (0.73/0.50 vs 0.71/0.46) and Churches (0.64/0.52 vs 0.61/0.44). On Bedrooms, precision is 0.66 vs 0.59/0.61, but recall ties StyleGAN at 0.48. 'Consistently improve' slightly overstates the result.

Second check: another model agreed

• p. 6 Figure 6 caption

  Results obtained with 100 DDIM steps [84] and κ = 0.

• p. 22 Figure 17 caption

  Results obtained with 100 DDIM steps [84] and κ = 0.

Only the authors can say what is right here.

Judge's note: Seen in the p.6 and p.22 images. A text search finds κ only in these two captions. η is probably intended, but only the authors can confirm.

Second check: another model agreed

• p. 25 Table 15, Class-Label-to-Image ImageNet, Model Size

  395M

• p. 7 Table 3, LDM-4 row

  400M

• p. 22 Table 10, LDM-4 (ours)

  400M

Only the authors can say what is right here.

Judge's note: p.25 Table 15 ImageNet column: f=4, 395M, batch 1200, 178K. p.22 Table 10 and p.7 Table 3: LDM-4 at 400M with 178K train steps and batch size 1200. The text layer also shows 400M in Table 18. 400M may be a rounded figure, but the other tables give sizes to the exact million, so the authors should say which value is right.

Second check: another model agreed

• p. 29 Sec. G, DM Training in Latent Space

  we estimate the component-wise variance

• p. 7 Sec. 4.3.2

  scaled by the component-wise standard deviation

Judge's note: Checked p.29 (formula, with μ̂ also summed over b,c,h,w) and the p.7 crop. Because the sums run over all indices, including channels, the estimate is one global scalar, not a per-component or per-channel variance. The formula is explicit, so the issue is that the wording 'component-wise' could mislead readers into expecting per-channel scaling.

Second check: another model agreed

• p. 24 Table 13, LDM-32 z-shape

  88 × 8 × 32

• p. 21 Table 8, f=32 VQ row

  32 16384 16 31.83

Only the authors can say what is right here.

Judge's note: Sec. 4.1 says 'Tab. 8 shows hyperparameters and reconstruction performance of the first stage models used for the LDMs compared in this section.' Every other LDM-f entry matches a Table 8 row: f=2/2048/c=2, f=4/8192/c=3, f=8/16384/c=4, f=16/16384/c=8. Only LDM-32 (c=32) has no matching row, since Table 8's f=32 VQ model has c=16. Only the authors can say which channel count is right.

Second check: another model agreed

• p. 22 Table 10, ImageBART/VQGAN+T rows and caption

  0.05 acc. rate∗

Only the authors can say what is right here.

Judge's note: p.22 image confirms both labels: the ImageBART and VQGAN+T rows say 'acc. rate' and the caption says 'rejection rate'. Read literally, the two labels give 0.05 vs 0.95 acceptance, so the authors should clarify which is meant.

Second check: another model agreed

• p. 28 Sec. F text

  we train all our models on a single NVIDIA A100 GPU

• p. 25 Table 15 caption

  except for the inpainting model which was trained on eight V100

Judge's note: Page 28 and page 25 images confirm both statements. In context, Sec. F is about the Table 18 models, which do not include inpainting, so this is unqualified wording rather than a contradiction in the data.

Second check: another model agreed

• p. 1 Figure 1 caption

  less agressive downsampling

• p. 3 Related work

  joint distributation

• p. 3 Related work

  partially adressed

• p. 7 Sec. 4.4

  diretly conditioning

Judge's note: Both misspellings are visible in the page images (p.1 Fig. 1 caption; p.7 Sec. 4.4). 'aggressive' is spelled correctly on p.3. The other misspellings grouped here, 'distributation' and 'adressed', are judged under C005 and C004.

Second check: none (judge only)

• p. 2 Contributions (vi)

  reusable for a various tasks besides training of DMs

• p. 2 Contribution (vi)

  reusable for a various tasks

Judge's note: Visible in the page 2 image, right column.

Second check: none (judge only)

• p. 2 Introduction, contribution (i)

  our method scales more graceful to higher dimen-

Judge's note: Visible in the page 2 image, contribution (i).

Second check: none (judge only)

• p. 3 Sec. 2, Diffusion paragraph

  We adress both drawbacks

• p. 3 Sec. 2, Diffusion Probabilistic Models paragraph

  the former can be partially adressed by advanced sampling

• p. 3 Sec. 2

  partially adressed by advanced sampling

Judge's note: Both are visible in the page 3 image, left column, Diffusion Probabilistic Models paragraph.

Second check: none (judge only)

• p. 3 Two-Stage Image Synthesis

  joint distributation

• p. 3

  learning a joint distributation over discretized image and text representations

Judge's note: Visible in the page 3 image, left column.

Second check: none (judge only)

• p. 3 3.1

  avoids bluriness

Judge's note: Visible in the page 3 image, right column, Sec. 3.1.

Second check: none (judge only)

• p. 3 Sec. 2

  the DM corresponds to a lossy compressor and allow to trade

Judge's note: Visible in the page 3 image, left column (also in crop C004-a). The subject is the singular 'the DM'.

Second check: none (judge only)

• p. 5 Sec. 4.2, first paragraph

  using ii) FID [29] and ii) Precision-and-Recall [50]

• p. 5 Sec. 4.2

  using ii) FID [29] and ii) Precision-and-Recall

Judge's note: Visible in the page 5 image, right column, Sec. 4.2.

Second check: none (judge only)

• p. 5 4.1

  in terms sampling speed

Judge's note: Visible in the page 5 image, right column.

Second check: none (judge only)

• p. 7 4.4 first sentence

  diretly conditioning on low-resolution images

• p. 7

  diretly conditioning on low-resolution images via concate-

Judge's note: The misspelling is visible in the p.7 page image, at the start of Sec. 4.4.

Second check: none (judge only)

• p. 7

  iffusion models. For text- to-image image modeling, we train

Judge's note: The p.7 page image shows "For **text-to-image** image modeling". The second "image" is redundant. This is a cosmetic slip.

Second check: code check

• p. 8 Table 5, third row

  emphLDM-4 (ours, big, 100 steps)

Judge's note: The label is visible in the Table 5 image on p.8.

Second check: none (judge only)

• p. 8 Sec. 4.5, first paragraph

  parts of the image are are corrupted

• p. 8

  er because parts of the image are are corrupted or to rep

Judge's note: The p.8 image shows "are" at the end of one line and again at the start of the next.

Second check: code check

• p. 11 References [33], [34]

  Image-to-image translation with conditional adver-

• p. 12 References [61] and [62]

  Semantic image synthesis with spatially-adaptive

• p. 11 References [40] and [41]

  A style-based generator architecture for generative adversarial networks.

Judge's note: In the p.11 and p.12 images, each pair names the same paper, venue and year in slightly different formats, and citations are split between the two numbers of each pair.

Second check: none (judge only)

• p. 20 Secs. D.1 and D.2

  the SNR is descreased

• p. 20 D.2

  autoenconding models

• p. 16 App. B

  typically choosen

Judge's note: I saw all three spellings in the page images: p.20 (D.1 and D.2) and p.16 (App. B).

Second check: none (judge only)

• p. 17 Eq. (13)

  KL(q(xt−1|xt, x0)|p(xt−1) =

Judge's note: The p.17 image shows the truncated expression. The earlier ELBO line (text layer line 1807) has KL(q(x_{t−1}|x_t,x_0)|p(x_{t−1}|x_t)). This is a typographical slip that does not change the derivation.

Second check: none (judge only)

• p. 27 Sec. E.3.6

  we followed the protocoll of [72] and and use the 2-alternative force-choice

• p. 27 E.3.6

  followed the protocoll of [72] and and use

• p. 20 D.1

  the SNR is descreased

• p. 23 D.6.1

  degration pipeline

Judge's note: Each one is visible in the page images: page 27 (E.3.6), page 20 (D.1) and page 23 (D.6.1).

Second check: code check

• p. 23 Sec. D.6.1

  seemed apppropiate for our application

• p. 23 D.6.1

  apppropiate

• p. 23 D.6.1

  bsr-degredation

• p. 23 D.6.1

  degration pipeline

Judge's note: The p.23 image shows all five misspellings in the D.6.1 paragraph.

Second check: none (judge only)

• p. 24 Table 13, LDM-32 z-shape

  88 × 8 × 32

• p. 25 Table 14, LDM-32 z-shape

  88 × 8 × 32

Judge's note: The p.24 (Table 13) and p.25 (Table 14) images both print '88 × 8 × 32'. The other columns halve each time (256, 128, 64, 32, 16), so the next value is 8, and 256/32 = 8.

Second check: none (judge only)

• p. 24 Table 12, row label and caption

  Iterations∗

Judge's note: The Table 12 caption on p.24 has no ∗ note. The Table 14 caption on p.25 does define ∗ ('All models are trained for 500k iterations...'), but Table 14 comes later, and its wording does not fit Table 12's varying iteration counts.

Second check: none (judge only)

• p. 27 Sec. E.3.1

  7.77 (Nichol and Dhariwal)

Judge's note: Page 26 image: 'script provided by Dhariwal and Nichol [15]'. Page 27 image: the order is reversed twice. The reference list in the text layer has [15] as 'Prafulla Dhariwal and Alex Nichol'. This is a cosmetic inconsistency in author order.

Second check: none (judge only)

• p. 27 Sec. E.3.3

  compute FID scores the 2048 unaugmented examples

Judge's note: Page 27 image confirms the wording.

Second check: none (judge only)

• p. 28 Table 18, FFHQ StyleGAN2 row

  32.13‡ - 32.13†

Judge's note: Page 28 crop: the markers are ‡ and † as described. The caption defines † as 'Numbers taken from [15]' and ‡ as 'Assumed to be trained on 25M train examples'. Both could apply, so this is only a marker inconsistency.

Second check: none (judge only)

• p. 28 App. F

  compute ressources

Judge's note: Seen on the p.28 page image. The same paragraph later spells it correctly: "required compute resources".

Second check: none (judge only)

• p. 29 Sec. H

  For our those models trained on comparably small datasets

• p. 29 App. H

  For our those models trained

Judge's note: Seen on the p.29 page image. Both 'our' and 'those' are printed.

Second check: none (judge only)