Writing study: True Self-Supervised Novel View Synthesis is Transferable

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Thomas MItchel, Hyunwoo Ryu, Vincent Sitzmann · 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) · Empirical analysis, diagnosis & auditing · Datasets, benchmarks & measurement resources.

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

17 physical pages

Before references / appendix

≈ 9.56 page-equivalents

Main figures / tables

1 / 3

Attached appendix figures / tables

2 / 6

Attached appendix span

≈ 4.7 page-equivalents

Before references / appendix: ≈ 56.24%References / intervening material: ≈ 16.12%Attached appendix: ≈ 27.65%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	≈ 1.06 pp.	6.24%
2 RELATED WORK	2	≈ 0.82 pp.	4.82%
3 METHOD	3	≈ 3.44 pp.	20.24%
4 EXPERIMENTS	6	≈ 2.98 pp.	17.53%
5 DISCUSSION	9	≈ 0.57 pp.	3.35%
ACKNOWLEDGMENTS	10	≈ 0.14 pp.	0.82%
REFERENCES	10	≈ 2.74 pp.	16.12%
A XFACTOR: ARCHITECTURE AND IMPLEMENTATION DETAILS	13	≈ 0.33 pp.	1.94%
B EVALUATION	13	≈ 4.37 pp.	25.71%

All detected figure and table locations

Main text

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

Attached appendix

• Table 4 · p. 14
• Table 5 · p. 14
• Table 6 · p. 15
• Table 7 · p. 15
• Figure 2 · p. 16
• Table 8 · p. 16
• Figure 3 · p. 17
• Table 9 · p. 17

Argument and organization

1. Define the task before the architecture

   Replace reconstruction-only evaluation with cross-scene trajectory transfer and formalize a diagnostic metric.

2. Two failure modes and XFactor

   Connect interpolation and information leakage to the stereo-monocular starting point and pose-preserving augmentation objective.

3. Transfer, probes, ablations and limitations

   Test the defining behavior separately from pose decodability and image reconstruction.

Problem, claim and contribution

The paper makes a testable behavioral criterion its central contribution: a pose latent should reproduce the same camera trajectory in another scene. It defines this before introducing the architecture, so the evaluation follows the task rather than a convenient reconstruction metric.

Sections 1 and 3.1–3.3, pp. 1–4. · Main paper

Argument and section order

Two named failure modes each lead to a concrete design choice. A one-context renderer removes interpolation, while complementary masks reduce target-pixel leakage; multi-view fine-tuning is then a separate training stage.

Sections 3.4–3.5, pp. 5–6. · Main paper

Figures, captions and visual explanation

Figure 1 uses paired photographs, complementary colored masks, two compact module labels and a rendering-loss endpoint. The same visual tokens carry the augmentation argument through the entire diagram without a wall of module descriptions.

Figure 1, p. 5. · Main paper

Tables, comparisons and evidence

Table 1 places rendered sequences and colored estimated trajectories directly above the quantitative comparison, coupling perceptual examples to the geometric test. Its caption states metric directions; Section 3.3 explicitly explains how TPS alone can be gamed.

Table 1, p. 7; Section 3.3, p. 4. · Main paper

Appendix and reproducibility

The appendix records baseline reimplementations, changed losses, training budgets, separate probe protocols and oracle failure rates. It also keeps ordinary autoencoding results separate from the claimed transfer task.

Appendices A–B, pp. 13–17; Tables 4–9. · Main paper

What to reuse—and what not to infer

Transferable lesson: State the behavior that would falsify the central claim, design the metric around it, and show why superficially similar reconstruction or probe scores are insufficient.

Boundary: Transferability is the authors' proposed operational criterion, not a proof that all previous NVS is invalid. VGGT supplies evaluation/probe poses, LPIPS remains a pretrained perceptual loss, and modified baselines have different training budgets. TPS requires a separate fidelity check, and Table 1 contains some suspicious nonmonotonic AUC entries outside the selected RE10K comparison.

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

Sections 1–5, pp. 1–10; references pp. 10–13.

Appendix / supplement

Appendices A–B, pp. 13–17: architecture, training and baseline differences, reconstruction, SSL-objective comparisons, temporal-corruption tests, oracle robustness and rendering artifacts.

Visually inspected PDF pages

5, 7

• ICLR 2026 official proceedings PDF, retrieved 2026-09-19 · 17 pages