Instant Neural Graphics Primitives with a Multiresolution Hash Encoding

• p. 2 Section 2, paragraph 2

  attention components of recurrent architectures [Gehring et al. 2017]

• p. 12 References

  Convolutional Sequence to Sequence Learning

• p. 2

  tion components of recurrent architectures [Gehring et al. 2017]

Only the authors can say what is right here.

Judge's note: Rests on outside knowledge: that the Gehring et al. ConvS2S model is non-recurrent; its title in the reference list (p.12) does say 'Convolutional'. The p.2 text attributes 'recurrent architectures' to this citation. Gehring et al. do use position embeddings and attention, so the citation is relevant to input encodings, but the descriptor 'recurrent' does not fit the cited work. Only the authors can say whether a different reference was meant or the wording should change.

Second check: another model agreed

• p. 8 Fig. 9 header

  Multiresolution hash encoding (Ours), 𝑇 = 15, 133 FPS

Judge's note: The crop of p.8 shows 'T = 15' with no base 2 or superscript, so this is not an extraction artefact. Everywhere else in the paper T is written as 2^k. The label presumably means T = 2^15.

Second check: another model agreed

• p. 13 References

  2020b. Convolutional Occupancy Networks. (2020). arXiv:2003.04618 [cs.CV]

• p. 11 Discussion, Generative setting

  typically a CNN such as StyleGAN [Chan et al. 2021; DeVries et al. 2021; Peng et al. 2020b]

• p. 11 Section 6, Generative setting

  [Chan et al. 2021; DeVries et al. 2021; Peng et al. 2020b]

• p. 13 Reference list

  2020b. Convolutional Occupancy Networks. (2020). arXiv:2003.04618

Judge's note: The p.13 crop shows both entries with identical authors and title. This is a duplicate entry. Citing it among CNN-populated feature-grid works is a loose fit rather than clearly wrong, since ConvONet fills a feature grid with a CNN. The defect is mainly the duplication; the citation could be merged into 2020a.

Second check: another model agreed

• p. 12 References, Lehmer 1951

  Harvard University Press, Cambridge, United Kingdom

• p. 12 Reference list, Lehmer 1951

  Harvard University Press, Cambridge, United Kingdom, 141–146.

Only the authors can say what is right here.

Judge's note: Rests on outside knowledge: Harvard University Press is located in Cambridge, Massachusetts, USA. The crop confirms the printed text.

Second check: another model agreed

• p. 1 Fig. 1 caption

  hyperparameters across all tasks and only vary the hash table size

• p. 2 Introduction

  configured by just two values—the number of parameters 𝑇and the desired finest resolution 𝑁max

• p. 4 Table 1 caption

  Only the hash table size𝑇and max. resolution 𝑁max need to be tuned to the task.

Judge's note: Checked the Fig. 1 caption, the Introduction and Table 1; Section 4 also sets Nmax per task. Nmax is set by a rule (2048 × scene size, half the image width, 2^19), so it is arguably not tuned. Even so, the caption's "only" conflicts with Table 1's own wording.

Second check: another model agreed

• p. 2 Introduction, third paragraph

  the number of parameters 𝑇

• p. 4 Section 3

  bounded by 𝑇· 𝐿· 𝐹which in our case is always 𝑇· 16 · 2

Judge's note: This is loose shorthand rather than an outright error, since the parameter count scales with T. It is still inconsistent with the formal definition.

Second check: another model agreed

• p. 3 Section 2, Sparse parametric encodings, last paragraph

  despite having 20× fewer parameters

• p. 3 Fig. 2 labels

  10 k + 494 k

• p. 3 Fig. 2 labels (d) and (e)

  10 k + 16.3 M

Only the authors can say what is right here.

Judge's note: Fig. 2 (p.3) counts: (c) 10 k + 33.6 M, (d) 10 k + 16.3 M, (e) 10 k + 494 k, (f) 10 k + 12.6 M. The sentence compares with the multiresolution grid (d). Against (d), (e) has about 32–33× fewer parameters and (f) about 1.3× fewer. Against (c), the ratios are about 67× and 2.7×. The 20× matches no printed pair. It may be a loose lower bound, but the paper does not say which configurations it refers to.

Second check: another model agreed

• p. 4 Table 1

  Max. entries per level (hash table size)

• p. 7 Fig. 6 caption

  With hash table sizes 𝑇 of 2

Judge's note: Table 1 and the Fig. 6 caption checked in the images. The stated range does not cover the 2^12 result.

Second check: another model agreed

• p. 6 Section 4, Performance considerations

  2 · 𝑇 · 𝐹 > 6 · 2

Only the authors can say what is right here.

Judge's note: With F = 2 the inequality holds only for T > 1.5·2^20. Several levels being resident in cache at once could explain the drop, but the text says "individual levels". The authors should clarify.

Second check: another model agreed

• p. 7 Section 5.2

  at roughly equal parameter count

• p. 8 Fig. 7 labels

  22.3 M (params)

Judge's note: I read the Fig. 7 labels on the p.8 image. Two NGLOD models have about 1.8× and 2.0× the hash parameter count. NGLOD's size depends on the octree for each shape, so the counts cannot be matched exactly. 'Roughly equal' is loose wording for up to a factor of 2.

Second check: another model agreed

• p. 9 Section 5.4, Model Architecture

  spherical harmonics basis (i.e. up to degree 4)

Judge's note: "Degree 4" may be meant as four bands, so the wording is ambiguous rather than clearly wrong.

Second check: another model agreed

• p. 2 Introduction, paragraph 2

  The important commonality of the these approaches

• p. 2

  The important commonality of the these approaches is an encod-

Judge's note: The p.2 image shows 'of the these approaches'. It should read 'of these approaches'.

Second check: none (judge only)

• p. 6 Section 4, Training

  as well a a batch size

Judge's note: Visible in the page image.

Second check: none (judge only)

• p. 6 Section 4, Training

  𝜖 = 10−15, The choice

• p. 6 Section 4, Training

  𝜖 = 10−15, The choice of 𝛽1 and 𝛽2 makes

Judge's note: Checked page 6. The comma is followed by a capitalised "The".

Second check: none (judge only)

• p. 9 Fig. 10 caption

  number of neurons the hidden layers of the density and color MLPs

Judge's note: Visible in the crop.

Second check: none (judge only)

• p. 10 Section 5.4, Comparison with high-quality offline NeRF

  the peak signal to noise ratio (PSNR) our NeRF implementation

Judge's note: Visible in the crop.

Second check: none (judge only)

• p. 12 References, Kingma and Ba

  arXiv:1412.6980 (June 2014)

• p. 12 Reference list

  arXiv:1412.6980 (June 2014).

Only the authors can say what is right here.

Judge's note: Rests on outside knowledge: arXiv identifier 1412 denotes a December 2014 submission. The crop confirms the printed text.

Second check: none (judge only)

• p. 12 References, Micikevicius et al.

  arXiv:1710.03740 (Oct. 2018)

• p. 12 Reference list

  arXiv:1710.03740 (Oct. 2018).

Only the authors can say what is right here.

Judge's note: Rests on outside knowledge: arXiv identifier 1710 denotes October 2017. The entry's year 2018 may reflect a later version, but the month and year do not fit together. The crop confirms the printed text.

Second check: none (judge only)

• p. 12 References (Gehring, Hadadan, Jansen, Nooruddin entries)

  1243—-1252

• p. 12 References (Nooruddin and Turk)

  191––205

Judge's note: The doubled dashes are visible in the page image, so this is not an extraction artefact. It is cosmetic.

Second check: none (judge only)

• p. 12 References, Nießner et al. and Nooruddin and Turk

  Article 169 (nov 2013)

• p. 12 References, Nooruddin and Turk

  (apr 2003)

Judge's note: Seen in the page image. This is a house-style inconsistency only.

Second check: none (judge only)

• p. 13 References, Pharr et al. 2016

  (3rd ed.) (3rd ed.)

• p. 13 Reference list

  From Theory to Implementation (3rd ed.) (3rd ed.)

Judge's note: Visible in the page image: once in the italic title and again as a separate edition field.

Second check: none (judge only)

• p. 14 Appendix E.1, Related work

  non-linear step size as benefitial

Judge's note: Seen in the crop of page 14.

Second check: none (judge only)