Writing study: Stable Video Infinity: Infinite-Length Video Generation with Error Recycling

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Wuyang Li, Wentao Pan, Po-Chien Luan, Yang Gao, Alexandre Alahi · 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: Learning & optimization algorithms (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

27 physical pages

Before references / appendix

≈ 10.2 page-equivalents

Main figures / tables

9 / 4

Attached appendix figures / tables

9 / 8

Attached appendix span

≈ 13.89 page-equivalents

Before references / appendix: ≈ 37.78%References / intervening material: ≈ 10.78%Attached appendix: ≈ 51.44%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	≈ 1.19 pp.	4.41%
2 RELATED WORK	2	≈ 0.53 pp.	1.96%
3 PRELIMINARIES AND MOTIVATION	3	≈ 1.3 pp.	4.81%
4 STABLE VIDEO INFINITY	4	≈ 3.12 pp.	11.56%
5 EXPERIMENTS	7	≈ 2.85 pp.	10.56%
6 CONCLUSION AND DISCUSSION	10	≈ 0.49 pp.	1.81%
ACKNOWLEDGEMENTS	11	≈ 0.09 pp.	0.33%
REFERENCES	11	≈ 2.91 pp.	10.78%
A BENCHMARK SETUP	14	≈ 1.51 pp.	5.59%
B DISCUSSION	15	≈ 1.76 pp.	6.52%
C QUANTITATIVE EXPERIMENTS	17	≈ 1.35 pp.	5%
D IMPLEMENTATION DETAILS	18	≈ 0.95 pp.	3.52%
E ADDITIONAL QUALITATIVE COMPARISON	19	≈ 8.32 pp.	30.81%

All detected figure and table locations

Main text

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

Attached appendix

• Figure 10 · p. 14
• Table 5 · p. 17
• Table 6 · p. 17
• Table 7 · p. 17
• Table 8 · p. 17
• Table 9 · p. 18
• Table 10 · p. 18
• Table 11 · p. 19
• Table 12 · p. 19
• Figure 11 · p. 20
• Figure 12 · p. 21
• Figure 13 · p. 22
• Figure 14 · p. 23
• Figure 15 · p. 24
• Figure 16 · p. 25
• Figure 17 · p. 26
• Figure 18 · p. 27

Argument and organization

1. Train–test mismatch

   Distinguishes single-clip prediction error from cross-clip conditioning error.

2. Closed-loop error recycling

   Explains injection, approximate error estimation, replay banking and the training objective.

3. Generation regimes and stress tests

   Separates consistent scenes, changing prompt streams and audio/skeleton conditions.

Problem, claim and contribution

The introduction reframes drifting as a training–inference mismatch and contrasts generation with restoration. This motivates training on errors rather than adding another inference-time heuristic.

Sections 1 and 3, pages 1–4 · Main paper

Argument and section order

The three error scenarios precede the replay policy, helping explain why image-conditioning errors are sampled differently from timestep-aligned latent errors.

Sections 4.2–4.3 and Figure 4, pages 5–6 · Main paper

Figures, captions and visual explanation

Figure 3 encodes the closed loop with numbered subpanels and distinct arrows for error sampling and updating. Latent, image and noise errors retain consistent labels across diagram and equations.

Figure 3, page 5 · Main paper

Tables, comparisons and evidence

Table 1 groups duration and scene-control regimes instead of pooling unlike tasks. Appendix C adds copied-frame and ping-pong baselines that expose how consistency metrics can reward trivial videos.

Table 1 page 7; Table 5 page 17 · Main paper

Appendix and reproducibility

The supplement makes reference-frame choices, LoRA settings, error probabilities and bank capacity explicit, then acknowledges style shifts, identity drift and lack of real-time streaming.

Appendices B.3 and D, pages 16 and 18–19 · Main paper

Clarity and writing conventions

The infinity framing is memorable but exceeds the finite evidence: principal benchmarks are 50/250 seconds and longer examples are demonstrations, not a guarantee of arbitrary-horizon coherence.

Section 5, pages 7–10 · Main paper

What to reuse—and what not to infer

Transferable lesson: For a feedback-based method, use one closed-loop figure and include an adversarially simple baseline that can game the headline metric.

Boundary: No finite evaluation proves infinite stable generation. Single-scene consistency and successful scene transitions are different objectives, and high consistency with near-zero motion can be a failure rather than a win.

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 1–6

Appendix / supplement

Pages 14–27, Appendices A–E including benchmark construction, limitations, ablations, settings and qualitative captions

Visually inspected PDF pages

5, 7

• ICLR 2026 final proceedings · 27 pages