Writing study: Generative Human Geometry Distribution

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Xiangjun Tang, Biao Zhang, Peter Wonka · Official ICLR 2026 oral record

Observed writing choices, not acceptance rules. This study examines the inspected manuscript version. It does not establish why the paper was accepted or that its presentation choices should be copied unchanged.

Figure, table and section ledger

Editorial contribution type: Architecture & representation design (primary).

Layout extraction, not a required paper template. Counts are distinct caption labels; a multi-panel figure counts once. Section spans are estimated from heading positions, including floats and whitespace. Method and limitations.

Canonical PDF

20 physical pages

Before references / appendix

≈ 9.19 page-equivalents

Main figures / tables

9 / 5

Attached appendix figures / tables

12 / 0

Attached appendix span

≈ 5.89 page-equivalents

Before references / appendix: ≈ 45.95%References / intervening material: ≈ 24.6%Attached appendix: ≈ 29.45%Estimated layout envelopes, not text/artwork area. Unresolved appendix boundaries remain within the undivided post-body material.

Section-by-section lengths and source pages

The title/abstract/front matter precede the first numbered heading. A span ends at the next detected top-level heading, not at the last sentence; end-page whitespace is included. References and statements are listed separately. These are not measurements of text-only or artwork area.

Top-level heading	Starts on PDF page	Estimated span	Share of entire PDF
1 INTRODUCTION	1	≈ 1.2 pp.	6%
2 RELATED WORK	2	≈ 1.1 pp.	5.5%
3 PRELIMINARIES	3	≈ 0.47 pp.	2.35%
4 METHOD	4	≈ 2.01 pp.	10.05%
5 EXPERIMENTS AND RESULTS	6	≈ 3.23 pp.	16.15%
6 LIMITATIONS	9	≈ 0.2 pp.	1%
7 CONCLUSIONS	9	≈ 0.38 pp.	1.9%
8 ACKNOWLEDGEMENTS	10	≈ 0.08 pp.	0.4%
REFERENCES	10	≈ 4.92 pp.	24.6%
9 APPENDIX	15	≈ 5.89 pp.	29.45%

All detected figure and table locations

Main text

• Figure 1 · p. 1
• Figure 2 · p. 2
• Table 1 · p. 3
• Figure 3 · p. 4
• Figure 4 · p. 5
• Figure 5 · p. 6
• Figure 6 · p. 7
• Figure 7 · p. 7
• Table 2 · p. 7
• Table 3 · p. 7
• Table 4 · p. 8
• Figure 8 · p. 8
• Table 5 · p. 8
• Figure 9 · p. 9

Attached appendix

• Figure 10 · p. 15
• Figure 11 · p. 15
• Figure 12 · p. 16
• Figure 13 · p. 16
• Figure 15 · p. 17
• Figure 14 · p. 17
• Figure 16 · p. 18
• Figure 17 · p. 18
• Figure 18 · p. 19
• Figure 20 · p. 19
• Figure 19 · p. 20
• Figure 21 · p. 20

Argument and organization

1. Representation bottleneck

   Frames memory and flow complexity as obstacles to scaling single-shape geometry distributions to a dataset.

2. Two-stage mechanism

   Defines source/target distributions, pairing and normalization before encoding a shape into a feature map and learning a second generative flow.

3. Component evidence before task evidence

   Places pairing and architecture ablations before geometry-distribution comparisons and the two downstream generation tasks.

Problem, claim and contribution

The introduction makes the representation the unit of novelty: storing each shape in network weights does not scale. Two named design changes then connect that bottleneck to a compact latent representation rather than merely announcing a new generative model.

Sections 1 and 4.1–4.3 · Main paper

Figures, captions and visual explanation

The overview uses two clearly separated panels for encoding and generation, with repeated feature-map colors, a visible SMPL condition and tensor dimensions beside intermediate objects. The comparison figure adds magnified crops to normal-map renderings so fine geometric differences can actually be inspected.

Figure 4, page 5; Figure 8, page 8 · Main paper

Tables, comparisons and evidence

Table 4 separates raw-geometry outputs from enhanced rendering, preventing post-processing from silently changing the comparison target. Table 2 also retains a single-shape setting where the proposed normalization is not the lowest-distance option, then explains why dataset-scale behavior differs.

Sections 5.2–5.3; Tables 2 and 4 · Main paper

Numerical headline claims need an independent arithmetic pass: the repeated 57% improvement does not equal the relative decrease from the displayed raw-geometry FID values 42.90 to 16.16, which is about 62.3%. The metric and exact values are safer to report than inheriting that percentage.

Abstract, Sections 1 and 5.3; Table 4 · Main paper

Appendix and reproducibility

The appendix expands the overview into module-level architecture, reports sampled point counts, GPU training time, inference memory and speed, and then groups extra outputs by the claim they support: challenging garments, body-shape variation and interpolation.

Section 9.1–9.5; Figures 10–19 · Main paper

Clarity and writing conventions

A dedicated limitations section names non-uniform sampling, unseen clothing styles and UV seams rather than offering a generic future-work sentence. The LLM-use statement separately describes language assistance and author responsibility; its position is an observed choice, not a mandatory 2026 heading.

Section 6; Section 9.6 · Main paper

What to reuse—and what not to infer

Transferable lesson: When introducing a representation, show both its internal data flow and the downstream artifact it improves. Align qualitative crops with the quantitative metric, and explicitly separate raw outputs from rendering enhancements.

Boundary: The two-task evidence is human-avatar generation, not general 3D reconstruction or pathology. The user study covers 25 participants and a small identity/pose set. Diagrammed optional conditioning modalities should not be mistaken for equally evaluated tasks, and headline percentages require arithmetic verification.

Apply this lesson using the argument and evidence worksheet.

Inspection record and sources

Checked

2026-09-19

Reading status

Full text + selected visuals inspected

Main text

Full abstract and Sections 1–8, pages 1–10, including formulation, two generation tasks, ablations, comparisons and limitations.

Appendix / supplement

Section 9.1–9.6 and Figures 10–21, pages 15–20: implementation, runtime, additional visual results, interpolation and LLM-use disclosure.

Visually inspected PDF pages

5, 8

• Official ICLR 2026 proceedings manuscript · 20 pages