Writing study: Locality-aware Parallel Decoding for Efficient Autoregressive Image Generation

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Zhuoyang Zhang, Luke Huang, Chengyue Wu, Shang Yang, Kelly Peng, Yao Lu, Song Han · 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

24 physical pages

Before references / appendix

≈ 10.11 page-equivalents

Main figures / tables

9 / 3

Attached appendix figures / tables

5 / 4

Attached appendix span

≈ 7.89 page-equivalents

Before references / appendix: ≈ 42.12%References / intervening material: ≈ 25%Attached appendix: ≈ 32.88%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	≈ 1.74 pp.	7.25%
2 METHOD	3	≈ 3.53 pp.	14.71%
3 EXPERIMENT	6	≈ 2.81 pp.	11.71%
4 ABLATION	9	≈ 0.66 pp.	2.75%
5 RELATED WORKS	10	≈ 0.51 pp.	2.13%
6 CONCLUSION	10	≈ 0.3 pp.	1.25%
REFERENCES	11	≈ 6 pp.	25%
APPENDIX	17	≈ 0.03 pp.	0.13%
A ADDITIONAL IMPLEMENTATION DETAILS	17	≈ 1.22 pp.	5.08%
B LOCALITY-AWARE GENERATION ORDER	18	≈ 2.5 pp.	10.42%
C MORE VISUALIZATION OF ATTENTION MAPS	20	≈ 1.25 pp.	5.21%
D MORE VISUALIZATION OF GENERATION EXAMPLES	22	≈ 2.09 pp.	8.71%
E EFFICIENCY ANALYSIS	24	≈ 0.8 pp.	3.33%

All detected figure and table locations

Main text

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

Attached appendix

• Table 4 · p. 17
• Table 5 · p. 17
• Table 6 · p. 20
• Figure 10 · p. 21
• Figure 11 · p. 21
• Figure 12 · p. 22
• Figure 13 · p. 23
• Table 7 · p. 24
• Figure 14 · p. 24

Argument and organization

1. Locality observation and locality-aware decoding

   Turns an empirical attention pattern into explicit proximity and repulsion rules for parallel token groups.

2. Decoupled queries, experiments, and implementation appendices

   Explains the training/inference contract before testing quality–latency trade-offs and exposing implementation choices.

Problem, claim and contribution

The paper distinguishes parallel prediction from merely selecting an order: query tokens expose target locations without pretending that simultaneously sampled image tokens are already known.

Sections 2.1–2.2; Figures 3–6, pp. 3–5 · Main paper

Argument and section order

Measured locality leads to two competing geometric rules, then order-construction pseudocode and spatial snapshots, then separate ablations of order and architecture.

Section 2.3, Algorithm 1 and Figure 8; Section 3.3 · Main paper

Figures, captions and visual explanation

The attention-mask diagrams use a shared color legend to contrast query visibility, image-token visibility, and cache contents. This makes a subtle architectural change inspectable beyond a prose list of modules.

Figures 4–6, pp. 4–5 · Main paper

Tables, comparisons and evidence

The main comparison places steps, parameters, FID, precision/recall, batch-one latency, and batch-64 throughput together. These columns prevent fewer steps from being mistaken for a hardware-independent speedup.

Table 1, p. 8; Section 3.1 · Main paper

Appendix and reproducibility

Implementation, order sensitivities, higher-resolution generation, and executable sampling logic are separated into appendices. However, the appendix swaps the names of the proximity/repulsion parameters relative to the main algorithm.

Appendices A–E; Appendix B.3 and Table 6, p. 20; Algorithm 1 · Main paper

Clarity and writing conventions

The raster-order counterpart is a particularly useful control, while claims about text-to-image scaling need the separate disclosure of the internal synthetic training corpus and recaptioning pipeline.

Tables 1 and 3; Appendix A.2 · Main paper

What to reuse—and what not to infer

Transferable lesson: Explain a parallel generator with a visibility-and-cache diagram, then report matched sequential and parallel variants with both quality and explicitly conditioned latency.

Boundary: This is a retrained architecture, not a lossless drop-in decoder. Same-group query interaction does not reveal sampled token values. Appendix parameter names, model sizes, and batch-size descriptions contain inconsistencies; consult the released configuration before reproducing them.

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

PDF pp. 1–10, Sections 1–4, including locality measurements, query-token training and decoding, order construction, and all main experiments.

Appendix / supplement

PDF pp. 17–24, Appendices A–E: implementation, ablations, text-to-image setup, qualitative examples, and sampling code.

Visually inspected PDF pages

4, 8, 20

• ICLR 2026 official proceedings PDF, retrieved 2026-09-19 · 24 pages