Photorealistic Text-to-Image Diffusion Models with Deep Language Understanding
Advances in Neural Information Processing Systems 35, 2022 · paper · AI and machine learning
Presents Imagen, a text-to-image diffusion model using a frozen pretrained text-only language model to encode prompts. It reports a COCO FID of 7.27, finds larger language models help most, and introduces the DrawBench benchmark, where raters preferred it.
Checked: PDF from arxiv.org, fetched 4 Oct 2026. Page numbers are those of that PDF.
11 findings
Substance The Introduction says "GLIDE [41] (at 12.4)", but Table 1 gives GLIDE's zero-shot FID-30K as 12.24, which rounds to 12.2.
- p. 3 Introduction
GLIDE [41] (at 12.4)
- p. 7 Table 1
12.24
Judge's note: p3 shows 12.4 and Table 1 on p7 shows 12.24. In the same sentence the other comparators match the table after rounding (DALL-E 2 10.39 to 10.4; Make-A-Scene 7.55 to 7.6). Only the GLIDE value does not.
Second check: another model agreed
Report a problem with this finding
Substance Sec. 4.2 says "Imagen achieves 39.2% preference rate" for photorealism on the original COCO set, but Table 2 (Original, Imagen) prints "39.5 ± 0.75%".
- p. 7 Sec. 4.2 text
Imagen achieves 39.2% preference rate
- p. 7 Table 2, Original/Imagen row
39.5 ± 0.75%
- p. 7
Imagen achieves 39.2% preference rate indicating high image quality generation
Only the authors can say what is right here.
Judge's note: Both values are on p7 and describe the same quantity: the original COCO validation set, Imagen, photorealism. Rounding cannot explain the 0.3-point gap, and no other analysis set is described. Only the authors can say which value is correct.
Second check: another model agreed
Report a problem with this finding
Substance Sec. 4.2 says "there is a boost in preference rate of Imagen to 43.6%" on the no-people set, but Table 2 (No people, Imagen) prints "43.9 ± 1.01%".
- p. 7 Sec. 4.2 text
boost in preference rate of Imagen to 43.6%
- p. 7 Table 2, No people/Imagen row
43.9 ± 1.01%
- p. 7 Sec. 4.2 text
there is a boost in preference rate of Imagen to 43.6%
- p. 7 Sec. 4.2
preference rate of Imagen to 43.6%
Only the authors can say what is right here.
Judge's note: Both values are on p7 and describe the same subset and metric. The gap is the same 0.3 points as in C007, which suggests the text quotes an earlier version of the table. Only the authors can say which value is correct.
Second check: another model agreed
Report a problem with this finding
Substance The App. F.1 and F.2 configs both give optax.warmup_cosine_decay_schedule(init_value=0.0, peak_value=1e-4, warmup_steps=10000, decay_steps=2500000, end_value=2500000); an end learning rate of 2.5e6 does not fit a 1e-4 peak.
- p. 44 App. F.1 config
end_value=2500000)
- p. 45 App. F.2 config
end_value=2500000)
Only the authors can say what is right here.
Judge's note: Rests on outside knowledge: in optax.warmup_cosine_decay_schedule, end_value is the final learning rate. Pages 44 and 45 both print end_value=2500000, the same number as decay_steps, which suggests a copy error. The prose says only '1e-4 learning rate' with 10000 warmup steps. The intended end value cannot be recovered from the files.
Second check: another model agreed
Report a problem with this finding
Reference or label Sec. 2.1 reads "BERT [15], T5 [51] and CLIP [46]". [51] is Raffel et al., "Online and Linear-Time Attention by Enforcing Monotonic Alignments", and [46] is Radford et al., "Learning to Generate Reviews and Discovering Sentiment". The paper cites T5 as [52] and CLIP as [49] elsewhere.
- p. 3 Sec. 2.1, second paragraph
BERT [15], T5 [51] and CLIP [46]
- p. 13 References [46], [51]
Online and Linear-Time Attention by Enforcing Monotonic Alignments
- p. 3 Sec. 2.1
T5 [51]
- p. 3 Sec. 2.1
T5 [51] and CLIP [46]
Judge's note: I checked the p3 image: it shows T5 [51] and CLIP [46], and two lines earlier the same page cites T5 [52]. On p13, [51] is the monotonic-attention paper, [52] is 'Exploring the Limits of Transfer Learning...' and [46] is the sentiment-reviews paper. App. D.1 (p22) has the same sentence with T5 [52] and CLIP [49], and Sec. 1 also cites CLIP [49].
Second check: another model agreed
Report a problem with this finding
Reference or label Reference [11] (PaLM) gives "In arXiv:2001.08361, 2022."; the arXiv identifier conflicts with the 2022 year and appears to belong to another paper.
- p. 11 Reference [11]
In arXiv:2001.08361, 2022.
- p. 11 Reference [11]
PaLM: Scaling Language Modeling with Pathways. In arXiv:2001.08361, 2022.
Only the authors can say what is right here.
Judge's note: Rests on outside knowledge: that arXiv identifiers starting 2001 date to January 2020, and that PaLM is arXiv:2204.02311. The entry as printed is confirmed on the page 11 image.
Second check: another model agreed
Report a problem with this finding
Reference or label Reference [16] gives "Diffusion models beat gans on image synthesis. In NeurIPS, 2022."; the venue year appears to be wrong.
- p. 11 Reference [16]
Diffusion models beat gans on image synthesis. In NeurIPS, 2022.
Only the authors can say what is right here.
Judge's note: Rests on outside knowledge: that Dhariwal & Nichol appeared at NeurIPS 2021. The entry as printed is confirmed on the page 11 image. Other NeurIPS 2021 entries in the same list ([13], [17]) are dated 2021.
Second check: another model agreed
Report a problem with this finding
Reference or label Table A.1 has footnote markers in "Ability to understand rare words3." and "DALLE-2 Reddit4.", but footnotes 3 and 4 have no text anywhere.
- p. 22 Table A.1, Rare Words and Reddit rows
Ability to understand rare words3.
- p. 21 Footnote
5https://github.com/openai/CLIP/blob/main/model-card.md
Only the authors can say what is right here.
Judge's note: On the page 22 image the table carries superscripts 3 and 4 but the page has no footnote area. Page 21 has only footnote 5 (CLIP model card). A text search finds no other footnote text. The intended content is unknown.
Second check: another model agreed
Report a problem with this finding
Reference or label The Fig. A.13(a) subcaption reads "Comparison between different text encoders.", but the panel plots Mean Pooling, Attention Pooling and Cross Attention, which App. D.3.1 describes as text-conditioning schemas.
- p. 28 Figure A.13(a) subcaption
(a) Comparison between different text encoders.
- p. 25 D.3.1
Fig. A.13a compares the CLIP-FID pareto curves for mean pooling,
- p. 25 Appendix D.3.1
Fig. A.13a compares the CLIP-FID pareto curves for mean pooling, attention pooling, and cross attention.
Judge's note: The page 28 image shows the legend entries Mean Pooling, Attention Pooling and Cross Attention. App. D.3.1 (page 25) cites Fig. A.13a for this conditioning comparison. The subcaption mislabels the panel.
Second check: another model agreed
Report a problem with this finding
Reference or label The Fig. A.29 caption for the UBlock reads 'number of ResNetBlock per DBlock numResNetBlocksPerBlock: int'; it should read 'per UBlock'.
- p. 42 Figure A.29 caption
number of ResNetBlock per DBlock
Judge's note: Page 42 image: the caption describes UBlock hyperparameters but says 'per DBlock'. The wording is identical to the Fig. A.28 DBlock caption on page 41, so it was evidently copied over.
Second check: another model agreed
Report a problem with this finding
Wording Sec. 4.3 says "pairwise comparison of Imagen with each of the three models", but four models are listed (DALL-E 2, GLIDE, Latent Diffusion, CLIP-guided VQ-GAN), and Fig. 3 has four panels.
- p. 7 Sec. 4.3
pairwise comparison of Imagen with each of the three models
- p. 8 Figure 3
Comparison between Imagen and DALL-E 2 [54], GLIDE [41], VQ-GAN+CLIP [12]
- p. 8 Figure 3 caption
Comparison between Imagen and DALL-E 2 [54], GLIDE [41], VQ-GAN+CLIP [12] and Latent Diffusion [57]
- p. 8 Fig. 3
VQ-GAN+CLIP [12]
Judge's note: The preceding sentence on p7 names four models. Fig. 3 on p8 has four panels (DALL-E 2, GLIDE, VQGAN+CLIP, Latent Diffusion), and its caption names all four.
Second check: another model agreed
Report a problem with this finding
19 minor (typos and the like; most checked by the judge only)
Reference or label Reference [45] reads "In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, 2022."; the conference year and the publication year disagree.
- p. 13 Reference [45]
Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, 2022.
Only the authors can say what is right here.
Judge's note: The entry is on the page 13 image and the inconsistency is visible within it. Only the authors (or outside sources) can say which year is right.
Second check: another model agreed
Report a problem with this finding
Wording The Fig. A.31(b) dynamic-thresholding pseudocode writes 's = jnp.max(s, 1.0)', where Sec. 2.3 describes thresholding only 'if s > 1', i.e. an element-wise maximum (jnp.maximum).
- p. 44 Figure A.31(b)
s = jnp.max(s, 1.0)
- p. 4 Sec. 2.3 Dynamic thresholding
if s > 1, then we threshold
Only the authors can say what is right here.
Judge's note: Rests on outside knowledge: in JAX, jnp.max(a, axis) takes an axis as its second argument and is a reduction; jnp.maximum is the element-wise maximum. Page 44 prints jnp.max(s, 1.0). Page 4 describes clamping s to at least 1. The figure is labelled pseudocode, so this is minor.
Second check: another model agreed
Report a problem with this finding
Wording Sec. 3 says "DrawBench comprises 200 prompts in total", while Appendix C says "11 categories with approximately 200 text prompts".
- p. 6 Sec. 3
DrawBench comprises 200 prompts in total
- p. 21 Appendix C
11 categories with approximately 200 text prompts
Judge's note: Both pages show these wordings. One gives an exact count and the other an approximate one. This is minor, but the exact prompt count is not settled.
Second check: another model agreed
Report a problem with this finding
Wording Sec. 4.4 says "human raters prefer T5-XXL over CLIP on DrawBench across all 11 categories" without saying which axis. App. D.1 limits this to "for image-text alignment", and in Fig. A.7(b) (fidelity) the T5-XXL share is below 50% for Conflicting, Marcus et al. and Reddit.
- p. 8 Sec. 4.4, 'Human raters prefer T5-XXL over CLIP on DrawBench'
human raters prefer T5-XXL over CLIP on DrawBench across all 11 categories
- p. 24 Figure A.7(b) Fidelity
(b) Fidelity
- p. 22 App. D.1
in all 11 categories for image-text alignment
- p. 8 Sec. 4.4, 'Human raters prefer T5-XXL over CLIP on DrawBench'
we find that human raters prefer T5-XXL over CLIP on DrawBench across all 11 categories
Judge's note: In Fig. A.7a (p24) T5-XXL is above 50% in all 11 categories. In Fig. A.7b I read T5-XXL at about 47% for Conflicting, 45% for Marcus et al. and 46% for Reddit. The main-text claim needs the alignment qualifier that App. D.1 gives.
Second check: another model agreed
Report a problem with this finding
Wording Fig. A.30 (Efficient U-Net for 64^2 -> 256^2) shows five DBlocks (256x, 128x, 64x, 32x, 16x) and five UBlocks, but the App. F.2 config for the same model lists four "blocks" entries, each with "strides": (2, 2).
- p. 43 Figure A.30
Figure A.30: Efficient U-Net architecture for 642 →2562.
- p. 45 App. F.2 architecture config
"blocks": [
Only the authors can say what is right here.
Judge's note: Pages 43 and 45 viewed. Four stride-2 blocks take 256 down to 16, which gives five resolutions, so the figure's labels may be resolutions rather than block entries. Neither the figure nor the text says how the five drawn blocks map to the four config entries, or which one has no stride.
Second check: another model agreed
Report a problem with this finding
Typo The adjusted x-prediction is printed as "(z_t − σ ε̃_θ)/α_t", where σ has no time subscript, unlike σ_t in the ε-prediction definition just below.
- p. 4 Sec. 2.2, before Eq. 2
(zt −σ˜ϵθ)/αt
Judge's note: The p4 image shows a plain σ in this expression and σ_t in the ε_θ definition that follows, with α_t used throughout. This is a notational slip.
Second check: none (judge only)
Report a problem with this finding
Typo Sec. 2.4: "We find this to be a critical for generating high fidelity images" has a stray article "a".
- p. 5 Sec. 2.4
We find this to be a critical for generating
Judge's note: The p5 image shows the phrase as quoted.
Second check: none (judge only)
Report a problem with this finding
Typo Sec. 4.3: "exceedingly prefer Imagen over all others models" should read "all other models".
- p. 7 Sec. 4.3
exceedingly prefer Imagen over all others models
Judge's note: The p7 image shows "all others models". The use of "exceedingly" is a style matter and is not counted.
Second check: none (judge only)
Report a problem with this finding
Typo Sec. 5 cites "text-to-image [57, 41, 57]", repeating reference 57 in one citation group.
- p. 8 Sec. 5 Related Work
text-to-image [57, 41, 57]
Judge's note: The p8 image shows [57, 41, 57]. It may be that a different reference was intended, but the text does not show which one.
Second check: none (judge only)
Report a problem with this finding
Typo The legends of Fig. 4(a) and Fig. A.5(a) read "T-Large" where "T5-Large" is meant; the other entries are T5-Small, T5-XL and T5-XXL. In addition, the Fig. 4(c) y-axis reads "FID@10K", while panels (a) and (b) read "FID-10K".
- p. 8 Figure 4(a) legend
T-Large
- p. 23 Figure A.5(a) legend
T-Large
- p. 8 Fig 4c y-axis
FID@10K
Judge's note: I saw "T-Large" in both legends (p8, p23) and the differing y-axis labels in Fig. 4.
Second check: none (judge only)
Report a problem with this finding
Typo Sec. 6 reads "all generative models, including Imagen, Imagen, may", with "Imagen" duplicated. Also, the App. D heading reads "Imagen Detailed Abalations and Analysis" (for Ablations).
- p. 9 Sec. 6, last paragraph
including Imagen, Imagen, may
- p. 9 Sec. 6
have have led researchers
- p. 21 App. D heading
Imagen Detailed Abalations and Analysis
Judge's note: The p9 image shows the duplicated word, and the p21 heading shows "Abalations". The "have have" mentioned in one note belongs to C015.
Second check: none (judge only)
Report a problem with this finding
Typo Sec. 6: "text-to-image models, which have have led researchers" repeats "have".
- p. 9 Sec. 6
have have led researchers
- p. 9 Sec. 6, third paragraph
which have have led researchers
- p. 9
f text-to-image models, which have have led researchers to
Judge's note: The p9 image shows "have have".
Second check: code check
Report a problem with this finding
Typo Sec. 6 says "A subset of this data was filtered to removed noise"; this should read "to remove".
- p. 9 Sec. 6
filtered to removed noise
- p. 9 Sec. 6, fourth paragraph
was filtered to removed noise
Judge's note: I saw the wording on the page 9 image (crop C016-a and the full page).
Second check: none (judge only)
Report a problem with this finding
Typo Sec. 6 says "We strongly caution against the use text-to-image generation methods"; the word 'of' is missing after 'use'.
- p. 9 Sec. 6
against the use text-to-image generation methods
- p. 9 Sec. 6, third paragraph
caution against the use text-to-image generation methods
Judge's note: I saw the wording on the page 9 image (crop C017-a).
Second check: none (judge only)
Report a problem with this finding
Typo Sec. 6 says "These considerations inform our decision to not to release code or a public demo"; the 'to' is doubled.
- p. 9 Sec. 6
our decision to not to release
- p. 9 Sec. 6, second paragraph
inform our decision to not to release
- p. 9 Sec. 6
decision to not to release
Judge's note: I saw the wording on the page 9 image (crop C018-a). The later paragraph uses the correct form "decision not to release".
Second check: none (judge only)
Report a problem with this finding
Typo AttnGAN (Xu et al., CVPR 2018) is listed twice, as [76] "AttnGAN: Fine-Grained Text to Image Generation..." and [77] "Attngan: Fine-grained text to image generation...".
- p. 15 References [76], [77]
Attngan: Fine-grained text to image generation with attentional generative adversarial networks.
- p. 15 References [76], [77]
Attngan: Fine-grained text to image generation with attentional generative adversarial networks
- p. 15 References [76], [77]
AttnGAN: Fine-Grained Text to Image Generation with Attentional Generative Adversarial
Judge's note: The page 15 image shows the same authors, title, venue and year in both entries.
Second check: none (judge only)
Report a problem with this finding
Typo The Appendix D heading reads "Imagen Detailed Abalations and Analysis"; 'Abalations' should be 'Ablations'.
- p. 21 Appendix D heading
Imagen Detailed Abalations and Analysis
Judge's note: I saw the heading on the page 21 image.
Second check: none (judge only)
Report a problem with this finding
Typo DrawBench category names differ between Table A.1 ("Colors", "Counting", "Description", "Rare Words") and the Fig. A.15 axis ("Color", "Count", "Descriptions", "Rarewords").
- p. 22 Table A.1
Description
- p. 29 Fig. A.15 axis
Descriptions
Judge's note: I compared the Table A.1 image on page 22 with the Fig. A.15 axis labels on page 29. The difference is cosmetic and either form is acceptable.
Second check: none (judge only)
Report a problem with this finding
Typo App. E says "they observe that DALLE-E 2 is worse than GLIDE"; this should read 'DALL-E 2'.
- p. 27 Appendix E
DALLE-E 2 is worse
Judge's note: I saw the wording on the page 27 image. The same paragraph elsewhere uses 'DALL-E 2'.
Second check: none (judge only)