Writing study: Déjà View: Looping Transformers for Multi-View 3D Reconstruction

NeurIPS overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← NeurIPS 2026 oral writing corpus

Alessandro Burzio, Tobias Fischer, Sven Elflein, Qunjie Zhou, Riccardo de Lutio, Jiawei Ren, Jiahui Huang, Shengyu Huang, Marc Pollefeys, Laura Leal-Taixé, Zan Gojcic, Haithem Turki · Official NeurIPS 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 (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

≈ 9.47 page-equivalents

Main figures / tables

4 / 5

Attached appendix figures / tables

5 / 3

Attached appendix span

≈ 4.91 page-equivalents

Before references / appendix: ≈ 47.35%References / intervening material: ≈ 28.1%Attached appendix: ≈ 24.55%Estimated layout envelopes, not text/artwork area. Unresolved appendix boundaries remain within the undivided post-body material.

Source-bound correction: Source-checked 2026-10-02: remove a bibliography title incorrectly interpreted as Appendix A on page 12. Supplementary Material for Deja View begins on physical page 16, followed by actual Appendices A-G; references occupy pages 10-15.

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	≈ 0.84 pp.	4.2%
2 Related Work	3	≈ 0.69 pp.	3.45%
3 Method	3	≈ 3.05 pp.	15.25%
4 Experiments	6	≈ 3 pp.	15%
5 Discussion	9	≈ 0.58 pp.	2.9%
Acknowledgements	10	≈ 0.06 pp.	0.3%
References	10	≈ 5.62 pp.	28.1%
Supplementary Material for Déjà View	16	≈ 0.04 pp.	0.2%
A Training Datasets	16	≈ 0.26 pp.	1.3%
B Two-Stage Depth Training	16	≈ 0.23 pp.	1.15%
C Emergent Iterative Correspondence Search	16	≈ 0.23 pp.	1.15%
D Scaling Beyond Kmax	16	≈ 0.79 pp.	3.95%
E Scaling Below Kmax	17	≈ 0.13 pp.	0.65%
F Baseline Configurations	17	≈ 1.83 pp.	9.15%
G Societal Impact	19	≈ 1.41 pp.	7.05%

All detected figure and table locations

Main text

• Figure 1 · p. 1
• Figure 2 · p. 4
• Figure 3 · p. 5
• Table 1 · p. 7
• Table 2 · p. 7
• Table 3 · p. 8
• Figure 4 · p. 9
• Table 4 · p. 9
• Table 5 · p. 10

Attached appendix

• Table 6 · p. 17
• Table 7 · p. 17
• Table 8 · p. 18
• Figure 5 · p. 18
• Figure 6 · p. 19
• Figure 7 · p. 19
• Figure 8 · p. 20
• Figure 9 · p. 20

Argument and organization

1. Problem and contribution

   The framing turns a broad depth-reduction goal into a testable hypothesis: some depth in multi-view reconstruction may serve iterative refinement. The architecture then shares a transformer block across iterations instead of presenting parameter reduction alone as an explanation.

2. Argument sequence

   The method moves from recurrent multi-view computation to a residual-stream diagnostic, then the training procedure. Evaluation proceeds from reconstruction/pose quality to resource costs and controlled variants; the appendix tests iteration counts beyond the trained regime.

3. Evidence and verification

   Tables 1–2 report geometric quality, Table 3 adds resource denominators, and the matched-data/iteration controls test whether the recurrent design itself explains the gains. The Table 3 caption limits the MASt3R FLOPs accounting to pair-network computation without global alignment, a necessary comparison qualification.

Problem, claim and contribution

The framing turns a broad depth-reduction goal into a testable hypothesis: some depth in multi-view reconstruction may serve iterative refinement. The architecture then shares a transformer block across iterations instead of presenting parameter reduction alone as an explanation.

Introduction pp. 1–2; hypothesis and architecture, §3.2, p. 4 · Main paper

Argument and section order

The method moves from recurrent multi-view computation to a residual-stream diagnostic, then the training procedure. Evaluation proceeds from reconstruction/pose quality to resource costs and controlled variants; the appendix tests iteration counts beyond the trained regime.

§§3.2–3.5, pp. 4–6; main evaluation pp. 7–10; Appendix D, p. 19 · Main paper

Figures, captions and visual explanation

Figure 2 preserves view colors from image tokens through alternating frame/global attention and marks the shared loop count K before separate camera/ray and depth decoders. Figure 3's direction-versus-norm analysis is distinct from that mechanism diagram. Table 3 aligns quality metrics with parameter count, FLOPs and peak memory under a stated multi-view setting.

Figure 2, p. 4; Figure 3, p. 5; Table 3, p. 8; rendered pages 4 and 8 · Main paper

Tables, comparisons and evidence

Tables 1–2 report geometric quality, Table 3 adds resource denominators, and the matched-data/iteration controls test whether the recurrent design itself explains the gains. The Table 3 caption limits the MASt3R FLOPs accounting to pair-network computation without global alignment, a necessary comparison qualification.

Tables 1–2, p. 7; Table 3, p. 8; Tables 4–5, pp. 9–10 · Main paper

Appendix and reproducibility

The appendix lists the multi-dataset training mixture and separates training-stage ablations from recurrence extrapolation. Figure 7 shows that adding untrained iterations can degrade performance; retaining this boundary is more useful than implying that a loop automatically enables unlimited test-time improvement.

Appendix A, pp. 16–17; Appendix B, p. 18; Appendices C–F and Figures 6–8, pp. 19–20 · Main paper

Clarity and writing conventions

The writing is strongest when it labels the iteration explanation as a hypothesis and distinguishes observed residual-direction stabilization from a fixed-point theorem. Resource claims remain interpretable when every quantity names the view count, device and accounting boundary.

§§3.2 and 3.4, pp. 4–5; Table 3 caption, p. 8; Appendix D, p. 19 · Main paper

What to reuse—and what not to infer

Transferable lesson: For a repeated/shared computation, pair a loop diagram with an explicit mechanism hypothesis, matched-training ablations, and a quality–resource table; test where extra iterations stop helping.

Boundary: Parameter, FLOP and peak-memory reductions are not automatically end-to-end wall-clock speedups. The public version's extrapolation experiment does not support indefinite performance scaling with recurrent steps.

Apply this lesson using the argument and evidence worksheet.

Inspection record and sources

Checked

2026-10-02

Reading status

Writing-focused manuscript scan + selected visuals inspected

Main text

Writing-oriented scan across manuscript framing, section sequence, figure/table captions, selected argument/evaluation prose and conclusion; not exhaustive technical or proof verification. Sections 1–5, pp. 1–10, including the table continuing onto p. 10; references pp. 10–15. Rendered the recurrent architecture p. 4 and resource/quality comparison p. 8. Source is the exact public author version listed below, not a verified accepted/camera-ready revision.

Appendix / supplement

Writing-oriented appendix/documentation scan: Appendices A–F, pp. 16–20: training mixture, two-stage training, attention variants, iteration extrapolation, early decoding and baseline implementation. Source code, data, model outputs and external video assets were not executed or independently validated.

Visually inspected PDF pages

4, 8

This is a writing-focused scan of the manuscript and the stated supplementary scope: framing, organization, visual explanation, evidence presentation and wording. It is not a line-by-line proof audit, reproduction of results or exhaustive scientific review.

• Public author arXiv manuscript 2605.30215v2; not verified identical to the accepted or camera-ready OpenReview revision · 20 pages