Writing study: AnyUp: Universal Feature Upsampling

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Thomas Wimmer, Prune Truong, Marie-Julie Rakotosaona, Michael Oechsle, Federico Tombari, Bernt Schiele, Jan Eric Lenssen · 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).

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

21 physical pages

Before references / appendix

≈ 10.11 page-equivalents

Main figures / tables

6 / 9

Attached appendix figures / tables

6 / 2

Attached appendix span

≈ 6.89 page-equivalents

Before references / appendix: ≈ 48.14%References / intervening material: ≈ 19.05%Attached appendix: ≈ 32.81%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	≈ 1.73 pp.	8.24%
2 RELATED WORK	3	≈ 0.41 pp.	1.95%
3 TASK FORMULATION	3	≈ 0.66 pp.	3.14%
4 LEARNING ENCODER-AGNOSTIC FEATURE UPSAMPLING	4	≈ 2.15 pp.	10.24%
5 EXPERIMENTS	6	≈ 3.77 pp.	17.95%
6 LIMITATIONS	10	≈ 0.24 pp.	1.14%
7 CONCLUSION	10	≈ 0.47 pp.	2.24%
REFERENCES	11	≈ 4 pp.	19.05%
A ADDITIONAL DETAILS ON RELATED WORK	15	≈ 0.38 pp.	1.81%
B ADDITIONAL EXPERIMENTS	15	≈ 0.84 pp.	4%
C IMPLEMENTATION DETAILS	16	≈ 0.77 pp.	3.67%
D COMPULSORY NOTE ON LLM USAGE	17	≈ 0.08 pp.	0.38%
E ADDITIONAL VISUALIZATIONS	17	≈ 4.81 pp.	22.9%

All detected figure and table locations

Main text

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

Attached appendix

• Table 10 · p. 15
• Table 11 · p. 16
• Figure 7 · p. 16
• Figure 8 · p. 17
• Figure 9 · p. 18
• Figure 10 · p. 19
• Figure 11 · p. 20
• Figure 12 · p. 21

Argument and organization

1. Encoder-agnostic feature upsampling

   Separate variable-channel compatibility from learned upsampling quality.

2. Architecture and crop supervision

   Explain canonical conditioning, unchanged value features and training losses.

3. Downstream quality and feature preservation

   Evaluate retrained probes separately from unchanged low-resolution probes.

Problem, claim and contribution

The introduction distinguishes an upsampler shared across encoders from a separate model trained for each encoder. Table 1 makes that distinction inspectable, while cross-encoder results remain empirical rather than proving universal accuracy.

§1 and Table 1, pp.1–3; Table 6, p.8 · Main paper

Argument and section order

The method progresses from variable-channel conditioning to local attention and then training consistency. Appendix Figure 8 directly illustrates the irrelevant distant attention that motivates the local window, supplying a visual explanation for an otherwise technical design choice.

§§4.1–4.3, pp.4–6; Figure 8, p.17 · Main paper

Figures, captions and visual explanation

Figure 3 separates crop-based supervision from inference and visibly routes original features into attention values. Figure 4 then expands only the unfamiliar canonicalization layer, illustrating shared filters, softmax and channel averaging instead of repeating the entire pipeline.

Figures 3–4, pp.4–5 · Main paper

Tables, comparisons and evidence

Tables 2–3 retrain dense-task probes, whereas Table 5 keeps existing low-resolution probes unchanged to test semantic preservation. Those are different claims: improvements after adaptation cannot alone establish an unchanged feature space.

§§5.1–5.2, Tables 2–5, pp.6–8 · Main paper

Appendix and reproducibility

The cost appendix specifies resolution and an A100. At 448² AnyUp takes 92.4 ms versus JAFAR's 113.9 ms, not a greater-than-50% runtime reduction; FeatUp is faster at 73.7 ms. Exact rows should replace a broad efficiency headline.

Appendix A, Table 10, p.15 · Main paper

Clarity and writing conventions

The limitations identify information that a linear combination of coarse features cannot recover. The implementation appendix also discloses a corrected JAFAR probe-training bug, an important qualification when comparing reproduced and originally published numbers.

§6, p.10; Appendix C, p.16 · Main paper

What to reuse—and what not to infer

Transferable lesson: Use separate evaluation protocols for task adaptation and representation preservation, and expand one decisive architectural operation in a second figure.

Boundary: Sharper RGB-guided features do not demonstrate recovery of unavailable high-resolution evidence. No WSI or cross-magnification pathology experiment is provided, and not every task or cost row improves.

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 1–7; references 11–14

Appendix / supplement

Pages 15–21, Appendices A–E, including costs, additional baselines, implementation and qualitative galleries

Visually inspected PDF pages

4, 5, 17, 21

• official ICLR 2026 proceedings PDF · 21 pages