Writing study: SANA-Video: Efficient Video Generation with Block Linear Diffusion Transformer

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Junsong Chen, Yuyang Zhao, Jincheng YU, Ruihang Chu, Junyu Chen, Shuai Yang, Xianbang Wang, Yicheng Pan, Zhou Daquan, Huan Ling, Haozhe Liu, Hongwei Yi, Hao Zhang, Muyang Li, Yukang Chen, Han Cai, Sanja Fidler, Ping Luo, Song Han, Enze Xie · 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) · Inference, algorithms & systems efficiency.

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

25 physical pages

Before references / appendix

≈ 10.11 page-equivalents

Main figures / tables

6 / 6

Attached appendix figures / tables

9 / 3

Attached appendix span

≈ 9.89 page-equivalents

Before references / appendix: ≈ 40.44%References / intervening material: ≈ 20%Attached appendix: ≈ 39.56%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	≈ 1.06 pp.	4.24%
2 PRELIMINARIES	3	≈ 0.37 pp.	1.48%
3 SANA-VIDEO	3	≈ 4.81 pp.	19.24%
4 EXPERIMENTS	8	≈ 1.38 pp.	5.52%
5 APPLICATIONS AND DEPLOYMENT	9	≈ 0.42 pp.	1.68%
6 RELATED WORKS	10	≈ 0.23 pp.	0.92%
7 CONCLUSION	10	≈ 0.65 pp.	2.6%
REFERENCES	11	≈ 5 pp.	20%
A LLM USAGE	16	≈ 0.12 pp.	0.48%
B FULL RELATED WORK	16	≈ 1.12 pp.	4.48%
C MORE IMPLEMENTATION DETAILS	17	≈ 1.4 pp.	5.6%
D MORE RESULTS	18	≈ 3.07 pp.	12.28%
E DATA PROCESSING PIPELINE	21	≈ 2.91 pp.	11.64%
F WORLD MODEL	24	≈ 1.27 pp.	5.08%

All detected figure and table locations

Main text

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

Attached appendix

• Table 7 · p. 18
• Figure 7 · p. 19
• Table 8 · p. 19
• Figure 8 · p. 20
• Figure 9 · p. 21
• Figure 10 · p. 21
• Figure 11 · p. 22
• Table 9 · p. 22
• Figure 12 · p. 23
• Figure 13 · p. 23
• Figure 14 · p. 24
• Figure 15 · p. 24

Separate source files—excluded from canonical-PDF totals

A separate author version may repeat main text and references. Whole-file lengths and counts are not extra appendix material and are never added to the canonical PDF totals. Read the study's source scope for which sections received substantive reading versus layout inspection.

Additional source 1

Separate PDF (or ZIP containing it): whole-file layout of 9 physical pages; 9 detected figures and 2 tables. These may include overlapping main-text material.

Detected heading	PDF page	Estimated span

No top-level headings detected; no section lengths inferred.

• Table 6 · p. 2
• Table 7 · p. 3
• Figure 7 · p. 4
• Figure 8 · p. 5
• Figure 9 · p. 5
• Figure 10 · p. 6
• Figure 11 · p. 6
• Figure 12 · p. 6
• Figure 13 · p. 7
• Figure 14 · p. 8
• Figure 15 · p. 8

Argument and organization

1. Efficiency bottlenecks lead to targeted components

   Moves from video token cost to linear attention, temporal locality, recurrent caches and training adaptation.

2. Different deployment regimes receive separate evidence

   Distinguishes short-video quality/latency, long-video training and consumer-GPU quantization.

Problem, claim and contribution

The opening connects high token counts to practical training and inference barriers, motivating architectural changes rather than treating parameter count as the only efficiency metric.

Introduction, pp. 1–3. · Main paper · Additional source 2

Argument and section order

The staged training account explains which capabilities are inherited from the image model and which require video and autoregressive adaptation.

Section 3.1, pp. 3–4; Appendix C.2, pp. 17–18. · Main paper · Additional source 2

Figures, captions and visual explanation

A three-panel overview links block ordering, the end-to-end pipeline and a magnified attention/FFN module; frozen and trained components are visually distinguished.

Figure 2, p. 4. · Main paper · Additional source 2

Tables, comparisons and evidence

The one-minute frame grid demonstrates a qualitative trajectory, while the neighboring quantitative table explicitly evaluates 30-second videos rather than implying the same test duration.

Figure 11 and Table 9, p. 22. · Main paper · Additional source 2

Appendix and reproducibility

Cache-update pseudocode and a staged compute/data breakdown supplement architecture dimensions, filtering criteria and caption prompts.

Appendix C, pp. 17–18; Appendix E, pp. 21–24. · Main paper · Additional source 2

Clarity and writing conventions

Named components provide navigation, but speed claims should always carry their baseline, resolution, precision and denoising regime; these vary across the paper.

Tables 2, 4 and 9; Figure 6. · Main paper · Additional source 2

What to reuse—and what not to infer

Transferable lesson: For a systems/model paper, give readers a hierarchy of pipeline → module → state update, and attach every efficiency claim to an explicit evaluation regime.

Boundary: Constant-size attention and convolution caches summarize history; they do not guarantee unchanged long-horizon quality or constant total output storage. Training-cost comparisons inherit pretrained image weights and different datasets. Video examples do not establish physical-world simulation fidelity or policy improvement.

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

Canonical paper pp. 1–10, Sections 1–7, including model, cache derivation, experiments and deployment.

Appendix / supplement

Canonical pp. 16–25, Appendices A–F; separately linked older 9-page supplementary appendix, printed pp. 15–23, also read in full.

Visually inspected PDF pages

4, 22

• ICLR 2026 official proceedings PDF, retrieved 2026-09-19 · 25 pages
• Older anonymous supplementary appendix, printed pages 15–23 (9 PDF pages), fully inspected; updated canonical appendix is embedded in the conference paper. · 9 pages