Writing study: MotionStream: Real-Time Video Generation with Interactive Motion Controls

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Joonghyuk Shin, Zhengqi Li, Richard Zhang, Jun-Yan Zhu, Jaesik Park, Eli Shechtman, Xun Huang · 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: Inference, algorithms & systems efficiency (primary) · Learning & optimization algorithms.

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

26 physical pages

Before references / appendix

≈ 10.11 page-equivalents

Main figures / tables

8 / 4

Attached appendix figures / tables

6 / 4

Attached appendix span

≈ 7.89 page-equivalents

Before references / appendix: ≈ 38.88%References / intervening material: ≈ 30.77%Attached appendix: ≈ 30.35%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	2	≈ 1 pp.	3.85%
2 RELATED WORK	3	≈ 0.67 pp.	2.58%
3 M O T I O N S T R E A M : STREAMING GENERATION MEETS MOTION CONTROLS	3	≈ 3.73 pp.	14.35%
4 EXPERIMENTS	7	≈ 3.18 pp.	12.23%
5 CONCLUSION	10	≈ 0.41 pp.	1.58%
REFERENCES	11	≈ 8 pp.	30.77%
A TRAINING EFFICIENT TINY VAE	19	≈ 0.76 pp.	2.92%
B VBENCH RESULTS AND USER STUDY	19	≈ 0.94 pp.	3.62%
C ADDITIONAL ABLATION EXPERIMENTS AND QUALITATIVE RESULTS	20	≈ 1.98 pp.	7.62%
D TRAINING DETAILS	22	≈ 2.36 pp.	9.08%
E EVALUATION PROTOCOLS	25	≈ 0.43 pp.	1.65%
F LIMITATION AND FUTURE WORK	25	≈ 1.05 pp.	4.04%
G ETHICS STATEMENT	26	≈ 0.36 pp.	1.38%

All detected figure and table locations

Main text

• Figure 1 · p. 1
• Figure 2 · p. 4
• Figure 3 · p. 5
• Table 1 · p. 7
• Figure 4 · p. 8
• Figure 5 · p. 8
• Table 2 · p. 8
• Table 3 · p. 8
• Figure 6 · p. 9
• Table 4 · p. 9
• Figure 7 · p. 10
• Figure 8 · p. 10

Attached appendix

• Table A1 · p. 19
• Table A2 · p. 19
• Table A3 · p. 20
• Figure A1 · p. 20
• Figure A2 · p. 21
• Figure A3 · p. 21
• Table A4 · p. 22
• Figure A4 · p. 23
• Figure A5 · p. 24
• Figure A6 · p. 26

Argument and organization

1. Motion-controlled teacher

   Introduces efficient conditioning before distillation.

2. Causal distillation

   Connects self-rollout, distribution matching and inference-time cache behavior.

3. Experiments and streaming demo

   Separates quality evaluation, streaming configuration and extra decoder optimization.

Problem, claim and contribution

Latency, causality and bounded duration are introduced as separate obstacles to interactive video. That framing supports a system contribution with several coordinated changes rather than pretending all gains arise from one loss term.

Introduction, p.2 · Main paper

Argument and section order

The training narrative explains why ordinary Self Forcing drifts beyond its horizon before describing sinks, rolling caches and cache-relative RoPE. It then shows exactly how the teacher's three-evaluation guidance is incorporated into one student evaluation per denoising step.

§3.2, pp.5–7 · Main paper

Figures, captions and visual explanation

Figure 2 separates teacher training and causal distillation with a clear stage boundary, reuses model colors and shows alternative attention masks. Inputs and conditioning paths remain visible rather than being hidden behind a single architecture box.

Figure 2, p.4 · Main paper

Tables, comparisons and evidence

Table 1 identifies different backbones and resolutions, while Appendix E discloses method-specific controls and resizing. The tiny-decoder experiment separately reports speed and quality, so the headline 29.5 FPS is not misattributed to the main full-VAE benchmarks.

Table 1, p.7; Tables A1–A2, p.19; Appendix E, p.25 · Main paper

Appendix and reproducibility

User preferences favor the larger ATI baseline for aesthetics even when track accuracy favors MotionStream. Explicit failure cases show scene-change and identity problems caused in part by retaining an initial-frame anchor.

Appendix B and Figure A1, p.20; Appendix F and Figure A6, pp.25–26 · Main paper

Clarity and writing conventions

Infinite-length means a bounded-cache streaming capability, not indefinitely verified quality. The quantitative extrapolation study reaches 241 frames, and claims of a perfect train–test match should be limited to the simulated cache mechanics.

§4.3, p.9; Appendix E–F, p.25 · Main paper

What to reuse—and what not to infer

Transferable lesson: For interactive systems, report latency and throughput separately, isolate infrastructure optimizations, and show how training reproduces the specific inference mechanism being claimed.

Boundary: Main quality results use the full VAE, not the 29.5-FPS tiny decoder. Three sampling steps remain; one NFE refers to one guided denoising evaluation. Baselines differ in backbone, conditioning and native resolution. Fixed sinks struggle with complete scene transitions, and motion-transfer evaluation is reconstruction-conditioned rather than unrestricted text-to-video generation.

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

pp.1–10, abstract and §§1–5; references pp.11–18

Appendix / supplement

pp.19–26, appendices A–G covering decoder, user study, ablations, training, evaluation, limitations and ethics

Visually inspected PDF pages

4, 19

• official ICLR 2026 proceedings PDF · 26 pages