Writing study: Depth Anything 3: Recovering the Visual Space from Any Views

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Haotong Lin, Sili Chen, Jun Hao Liew, Donny Y. Chen, Zhenyu Li, Yang Zhao, Sida Peng, Hengkai Guo, Xiaowei Zhou, Guang Shi, Jiashi Feng, Bingyi Kang · 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

25 physical pages

Before references / appendix

≈ 10.11 page-equivalents

Main figures / tables

6 / 6

Attached appendix figures / tables

3 / 8

Attached appendix span

≈ 8.89 page-equivalents

Before references / appendix: ≈ 40.44%References / intervening material: ≈ 24%Attached appendix: ≈ 35.56%Estimated layout envelopes, not text/artwork area. Unresolved appendix boundaries remain within the undivided post-body material.

Source-bound correction: Depth Anything 3: Figure 7 is the real qualitative poor-depth-label panel on physical page 18. Poppler emitted Figure and 7: as separate same-baseline lines. Visually checked page 18 and restored the caption.

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	2	≈ 1.32 pp.	5.28%
2 DEPTH ANYTHING 3	3	≈ 2.05 pp.	8.2%
3 VISUAL GEOMETRY BENCHMARK	5	≈ 0.4 pp.	1.6%
4 EXPERIMENTS	5	≈ 3.9 pp.	15.6%
5 RELATED WORK	9	≈ 0.87 pp.	3.48%
6 CONCLUSION AND DISCUSSION	10	≈ 0.46 pp.	1.84%
REFERENCES	11	≈ 6 pp.	24%
A METHOD DETAILS	17	≈ 2.54 pp.	10.16%
B VISUAL GEOMETRY BENCHMARK	19	≈ 1.6 pp.	6.4%
C EXPERIMENTAL SETUP	21	≈ 1.45 pp.	5.8%
D ADDITIONAL ANALYSIS	22	≈ 0.07 pp.	0.28%
E ADDITIONAL EXPERIMENTS	22	≈ 3.24 pp.	12.96%

All detected figure and table locations

Main text

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

Attached appendix

• Table 7 · p. 18
• Table 8 · p. 18
• Figure 7 · p. 18
• Table 9 · p. 21
• Table 10 · p. 22
• Table 11 · p. 22
• Table 12 · p. 23
• Table 13 · p. 23
• Figure 8 · p. 23
• Table 14 · p. 24
• Figure 9 · p. 25

Argument and organization

1. Depth Anything 3

   Move from geometric representation to architecture, teacher supervision and the downstream rendering extension.

2. Visual geometry benchmark

   Define three evaluation targets before presenting comparisons.

3. Sufficiency ablations

   Test target choices and transformer organization rather than relying only on leaderboard results.

Problem, claim and contribution

Two questions—what to predict and how much architecture is necessary—organize the introduction. The contribution is framed around removing task-specific redundancy, then tested through separate target and backbone ablations.

§1, p.2; §§4.2–4.3, p.9 · Main paper

Argument and section order

Representation precedes implementation: depth and unnormalized rays first explain point recovery, followed by token attention, camera conditioning and the dual head. The rendering extension is then identifiable as a separate fine-tuning task.

§§2.1–2.4, pp.3–5 · Main paper

Figures, captions and visual explanation

The pipeline distinguishes optional camera inputs, shared feature extraction, prediction heads and supervision with different arrows. The teacher-label comparison separately shows thin-structure detail; it is qualitative support, not evidence that every geometric metric improves.

Fig.2, p.3; Fig.8, p.23 · Main paper

Tables, comparisons and evidence

Pose-free and pose-conditioned reconstruction occupy separate columns, while the paper explicitly acknowledges that ScanNet++ is not dataset-level zero-shot. This protects against attributing supplied camera information or familiar data domains to universal generalization.

§3, p.5; Table 2, p.7; Appendix C.1, p.21 · Main paper

Appendix and reproducibility

Teacher construction, alignment, training-stage transition and benchmark thresholds are indispensable details placed in the appendix. Table 7 also exposes the speed–accuracy trade-off between the auxiliary camera head and extracting poses from rays.

Appendices A–C, pp.17–22 · Main paper

Clarity and writing conventions

Some categorical prose is stronger than its tables: the extra point head slightly improves the reported averages, and removing teacher labels improves selected metrics. A careful summary should preserve those exceptions rather than copy the sufficiency rhetoric wholesale.

Table 5, p.8; §4.2, p.9; Table 14, p.24 · Main paper

What to reuse—and what not to infer

Transferable lesson: Let a compact representation explain the design, then separate benchmark results, sufficiency tests and implementation trade-offs into different evidence blocks.

Boundary: Minimal does not mean unmodified execution or one loss: cross-view token grouping, dual fusion branches, camera tokens and multiple supervised terms remain. Pose extraction speed is not end-to-end speed. HiRoom's scene count differs between appendix prose and Table 9. PixelSplat receives fewer NVS training steps, and several ablation claims have table-level exceptions.

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

PDF pp.1–10, §§1–6; references pp.11–16

Appendix / supplement

Appendices A–E and disclosure, PDF pp.17–25, including teacher design, benchmark construction, all ablations and qualitative captions

Visually inspected PDF pages

3, 23, 24

• official ICLR 2026 proceedings PDF · 25 pages