Writing study: WAFT: Warping-Alone Field Transforms for Optical Flow

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Yihan Wang, Jia Deng · 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

14 physical pages

Before references / appendix

≈ 9.27 page-equivalents

Main figures / tables

4 / 6

Attached appendix figures / tables

0 / 1

Attached appendix span

≈ 0.21 page-equivalents

Before references / appendix: ≈ 66.21%References / intervening material: ≈ 32.29%Attached appendix: ≈ 1.5%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	≈ 1.13 pp.	8.07%
2 RELATED WORK	2	≈ 1.18 pp.	8.43%
3 BACKGROUND	3	≈ 1.98 pp.	14.14%
4 METHOD	5	≈ 0.75 pp.	5.36%
5 EXPERIMENTS	6	≈ 3.69 pp.	26.36%
ACKNOWLEDGMENTS	10	≈ 0.06 pp.	0.43%
REFERENCES	10	≈ 4.53 pp.	32.36%
A APPENDIX	14	≈ 0.21 pp.	1.5%

All detected figure and table locations

Main text

• Figure 1 · p. 2
• Figure 2 · p. 3
• Figure 3 · p. 4
• Table 1 · p. 4
• Figure 4 · p. 5
• Table 3 · p. 6
• Table 2 · p. 7
• Table 4 · p. 8
• Table 5 · p. 9
• Table 6 · p. 14

Attached appendix

• Table 6 · p. 14

Argument and organization

1. Cost-volume necessity as the question

   Challenge a standard operation rather than claim an entirely new task.

2. Background before simplification

   Define cost volumes and warping so that removing one has a clear meaning.

3. Benchmark, transfer and component tests

   Separate fine-tuned rankings, zero-shot generalization and mechanism ablations.

Problem, claim and contribution

The central question is unusually crisp: must accurate optical flow explicitly construct a cost volume? The claimed simplification retains recurrent residual updates and a transformer, so the contribution is removal of particular machinery, not absence of inductive design.

§1, pp.1–2; §4, pp.5–6 · Main paper

Argument and section order

A background section explicitly defines the costly representation before the method removes it. Separate descriptions of frozen backbones and trainable adaptation prevent readers from mistaking pretrained feature use for training the entire backbone.

§§3.1–3.3 and §4, pp.3–6 · Main paper

Figures, captions and visual explanation

Figure 2 holds the two pixel grids fixed while changing all-pairs, neighborhood and single-location connections. This visual comparison communicates the memory argument before algebra. Figure 1 then shows the single warping operation inside the recurrent loop.

Figures 1–2, pp.2–3 · Main paper

Tables, comparisons and evidence

The experiments distinguish leaderboard fine-tuning from Chairs/Things-only transfer. Table 5 tests warping, resolution, recurrent architecture and context features; its CNN and no-warp variants support the mechanism more directly than cross-paper rankings do.

§§5.2–5.5, Tables 2–5, pp.7–9 · Main paper

Appendix and reproducibility

Sequence-level results expose a serious Ambush 1 failure: Twins-a2 scores 20.561 final EPE versus Flowformer++'s 6.610. Excluding that sequence changes the comparison but must supplement, not replace, the complete benchmark.

Appendix A, Table 6, p.14 · Main paper

Clarity and writing conventions

The abstract's broad first-place wording needs narrowing: Table 2 has DDVM's 3.26 KITTI-all versus DAv2-a2's 3.31, and Flowformer++ beats WAFT on Sintel-final. Table 6's SAMFlow average also appears to reverse clean and final ordering.

Abstract; Tables 2 and 6, pp.7,14 · Main paper

What to reuse—and what not to infer

Transferable lesson: Illustrate the expensive operation and its replacement with matched diagrams, then distinguish mechanism ablations from benchmark and transfer evidence.

Boundary: Preserve all benchmark sequences, backbone variants and timing hardware. Residual flow updates are not residual knowledge distillation, and lower memory does not mean fastest inference against every baseline.

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

Pages 1–10, Sections 1–5.5; references 10–14

Appendix / supplement

Appendix A and complete sequence-wise Table 6, p.14

Visually inspected PDF pages

2, 3, 14

• official ICLR 2026 proceedings PDF · 14 pages