Writing study: TRACE: Your Diffusion Model is Secretly an Instance Edge Detector

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Sanghyun Jo, Ziseok Lee, Wooyeol Lee, Jonghyun Choi, Jaesik Park, Kyungsu Kim · 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) · Architecture & representation design · Empirical analysis, diagnosis & auditing.

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

39 physical pages

Before references / appendix

≈ 10.28 page-equivalents

Main figures / tables

11 / 7

Attached appendix figures / tables

10 / 9

Attached appendix span

≈ 21.9 page-equivalents

Before references / appendix: ≈ 26.36%References / intervening material: ≈ 17.49%Attached appendix: ≈ 56.15%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	≈ 1.55 pp.	3.97%
2 RELATED WORK	3	≈ 1.52 pp.	3.9%
3 METHOD	4	≈ 2.02 pp.	5.18%
4 EXPERIMENTS	6	≈ 0.82 pp.	2.1%
5 ABLATION STUDY	7	≈ 1.17 pp.	3%
6 DISCUSSION	8	≈ 2.12 pp.	5.44%
7 CONCLUSION	10	≈ 0.36 pp.	0.92%
8 ACKNOWLEDGMENTS	11	≈ 0.17 pp.	0.44%
REFERENCES	11	≈ 6.83 pp.	17.51%
Appendix	18	≈ 1 pp.	2.56%
A LLM USAGE DISCLOSURE	19	≈ 0.12 pp.	0.31%
B COMPREHENSIVE REVIEW OF RELATED WORK	19	≈ 4.39 pp.	11.26%
C METHOD DETAILS	23	≈ 3.01 pp.	7.72%
D INFORMATION THEORETIC VIEW OF IEP	26	≈ 1.6 pp.	4.1%
E COMPREHENSIVE ANALYSIS AND DISCUSSION	28	≈ 4.2 pp.	10.77%
F DATASETS AND METRICS FOR EVALUATION	32	≈ 7.58 pp.	19.44%

All detected figure and table locations

Main text

• Figure 1 · p. 2
• Figure 2 · p. 2
• Figure 3 · p. 3
• Figure 4 · p. 4
• Figure 5 · p. 5
• Figure 6 · p. 6
• Table 1 · p. 7
• Table 2 · p. 7
• Table 3 · p. 7
• Figure 7 · p. 8
• Figure 8 · p. 8
• Table 4 · p. 8
• Table 5 · p. 8
• Figure 9 · p. 9
• Figure 10 · p. 9
• Figure 11 · p. 9
• Table 6 · p. 10
• Table 7 · p. 10

Attached appendix

• Table 8 · p. 21
• Figure 12 · p. 22
• Figure 13 · p. 23
• Table 9 · p. 28
• Table 10 · p. 29
• Figure 14 · p. 29
• Table 11 · p. 29
• Table 12 · p. 30
• Table 13 · p. 31
• Table 14 · p. 31
• Table 15 · p. 31
• Figure 15 · p. 32
• Table 16 · p. 33
• Figure 16 · p. 34
• Figure 17 · p. 35
• Figure 18 · p. 36
• Figure 19 · p. 37
• Figure 20 · p. 38
• Figure 21 · p. 39

Argument and organization

1. Observed transient instance cues

   Derives the method from a specific attention phenomenon.

2. IEP, ABDiv, distillation and propagation

   Assigns each stage a distinct estimation or efficiency job.

3. Benefits and domain failures

   Tests downstream usefulness without hiding negative transfer.

Problem, claim and contribution

Two visible errors—merging adjacent objects and fragmenting single objects—organize the motivation. The proposed boundary cue answers those errors directly rather than treating generative features as generally superior representations.

Figures1–3 and §1, pp.1–3 · Main paper

Argument and section order

The method names four operations and the appendix maps each to the two motivating failures. Algorithms expose uncertainty masking and the final downstream propagation, avoiding a diagram that hides essential post-processing.

§§3.2–3.5, pp.5–6; Table8 and Algorithms1–4, pp.21–26 · Main paper

Figures, captions and visual explanation

Figure4 separates frozen pseudo-edge generation from trainable one-step distillation using freeze/flame symbols, then shows edge gaps closing during training. The later galleries display edges alongside both baseline and refined masks, exposing the intermediate output.

Figure4, p.4; Figures17–21, pp.35–39 · Main paper

Tables, comparisons and evidence

Table1 separates pseudo-mask inference from downstream segmentation fine-tuning, but training-free should not describe the entire system: TRACE itself is LoRA-trained on ImageNet. The 81× figure compares its one-pass predictor with its own multi-step edge extraction, not complete segmentation pipelines.

Tables1,3,7, pp.7–10; AppendixE.3, p.30 · Main paper

Appendix and reproducibility

The strongest qualification is quantitative negative transfer: COIN's MoNuSeg PQ falls from0.536 to0.439 after TRACE, and TNBC from0.540 to0.426. The failure figures visibly show mismatched cellular boundaries, turning domain limitations into evidence.

AppendixE.4, Table15 and Figure15, pp.31–32 · Main paper

Clarity and writing conventions

The information-theoretic explanation should remain an interpretation: Eq.16 contains squared schedule increments, so maximizing KL need not maximize Fisher information alone. Table14 also labels65.4→50.3 as−5.1 rather than−15.1, illustrating why displayed deltas require independent checking.

AppendixD.2, pp.27–28; Table14, p.31 · Main paper

What to reuse—and what not to infer

Transferable lesson: Show the intermediate representation, explain what each pipeline stage repairs, and include numerical negative-transfer cases alongside attractive qualitative successes.

Boundary: Annotation-free means no new instance labels for edge learning, not untrained or free of pretrained-model supervision. Claims about universal diffusion advantages, tiny objects and pathology must be limited by the tested backbones and documented failures.

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.1–11, §§1–8; references pp.11–17

Appendix / supplement

pp.18–39, A–F including all algorithms, information-theoretic interpretation, failures, metrics and qualitative galleries

Visually inspected PDF pages

4, 31, 32, 34, 35, 36, 37, 38, 39

• official ICLR 2026 proceedings PDF · 39 pages