Writing study: Motion Attribution for Video Generation

ICML overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICML 2026 oral writing corpus

Xindi Wu, Despoina Paschalidou, Jun Gao, Antonio Torralba, Laura Leal-Taixé, Olga Russakovsky, Sanja Fidler, Jonathan Lorraine · Official ICML 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: Training, objectives & optimization (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

22 physical pages

Before references / appendix

≈ 8.19 page-equivalents

Main figures / tables

5 / 2

Attached appendix figures / tables

4 / 5

Attached appendix span

≈ 9.96 page-equivalents

Before references / appendix: ≈ 37.23%References / intervening material: ≈ 17.5%Attached appendix: ≈ 45.27%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	≈ 0.85 pp.	3.86%
2. Background	2	≈ 1.06 pp.	4.82%
3. Method	3	≈ 2.64 pp.	12%
4. Experiment	5	≈ 3 pp.	13.64%
5. Conclusion	8	≈ 0.17 pp.	0.77%
Acknowledgements	9	≈ 0.09 pp.	0.41%
Impact Statement	9	≈ 0.06 pp.	0.27%
References	9	≈ 3.85 pp.	17.5%
A. Notation	13	≈ 2.04 pp.	9.27%
B. Extended Related Work	15	≈ 0.74 pp.	3.36%
C. Additional Experiments	15	≈ 0.73 pp.	3.32%
D. Analysis	16	≈ 1.12 pp.	5.09%
E. Additional Method Details	17	≈ 0.72 pp.	3.27%
F. Additional Experiment Details	18	≈ 2 pp.	9.09%
G. Discussion	20	≈ 0.92 pp.	4.18%
H. Visualization	21	≈ 1.69 pp.	7.68%

All detected figure and table locations

Main text

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

Attached appendix

• Table 3 · p. 13
• Table 4 · p. 14
• Table 5 · p. 16
• Figure 6 · p. 16
• Figure 7 · p. 17
• Figure 8 · p. 19
• Table 6 · p. 20
• Table 7 · p. 20
• Figure 9 · p. 22

Argument and organization

1. Problem and contribution

   Identifies compute, variable clip duration and temporal failure modes as distinct obstacles to transferring image attribution into video.

2. Argument sequence

   Method subsections separately solve estimator scale, duration bias and motion emphasis before selection and explicit computational accounting.

3. Evidence and verification

   Human comparisons state 50 videos and 17 participants, while projection analysis tests agreement with full gradients rather than treating compression as exact.

Problem, claim and contribution

Identifies compute, variable clip duration and temporal failure modes as distinct obstacles to transferring image attribution into video.

Contributions and background, p. 2 · Main paper

Argument and section order

Method subsections separately solve estimator scale, duration bias and motion emphasis before selection and explicit computational accounting.

Sections 3.2-3.6, pp. 3-5 · Main paper

Figures, captions and visual explanation

The results page combines a normalization example, a projection-dimension tradeoff curve and a human preference table that retains ties and losses; however the normalization caption reverses the panels' visible with/without labels.

Figures 4-5 and Table 2, p. 8; PNG inspected; background p. 2 also visually inspected · Main paper

Tables, comparisons and evidence

Human comparisons state 50 videos and 17 participants, while projection analysis tests agreement with full gradients rather than treating compression as exact.

Sections 4.3-4.4, pp. 7-8 · Main paper

Appendix and reproducibility

Runtime tables separate an approximately 150-hour one-time gradient pass from about 54-second new-query work, and limitations retain camera-motion and guidance mismatches.

Appendix G and Tables 6-7, pp. 20-21 · Main paper

Clarity and writing conventions

The appendix candidly distinguishes amortized usefulness from higher upfront cost; motion-overlay captions call the visualization heuristic intuition rather than causal validation.

Appendix G.1, p. 20; Figure 9, p. 22 · Main paper

What to reuse—and what not to infer

Transferable lesson: For amortized selection methods, report both initialization cost and per-query cost, and keep preference ties and approximation error visible.

Boundary: Motion attribution is an approximate estimator, not exact training-data causality; retain the 10k-sample experiment scope and resolve the Figure 4 left/right caption mismatch before reuse.

Apply this lesson using the argument and evidence worksheet.

Inspection record and sources

Checked

2026-10-07

Reading status

Writing-focused manuscript scan + selected visuals inspected

Main text

Writing-focused distributed manuscript scan, not exhaustive technical/proof validation. Sections 1-5, pp. 1-8: video-specific attribution obstacles, estimator, duration correction, motion weighting, selection, costs and ablations.

Appendix / supplement

Writing-focused appendix scan; code, data and external assets not executed. Appendices A-H, pp. 13-22: notation, related work, additional model results, motion analysis, method/setup, runtime limitations and overlays.

Visually inspected PDF pages

2, 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.

• Published ICML 2026 proceedings PDF, PMLR volume 306; retrieved 2026-10-07. Public OpenReview decision notes were not independently reverified. · 22 pages