Writing study: Orthogonal Concept Erasure for Diffusion Models

ICML overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICML 2026 oral writing corpus

Yuhao Sun, Lingyun Yu, Hao-Xiang Xu, Fengyuan Miao, Zhuoer Xu, Hongtao Xie · Official ICML 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: Training, objectives & optimization (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

22 physical pages

Before references / appendix

≈ 9.21 page-equivalents

Main figures / tables

4 / 6

Attached appendix figures / tables

4 / 7

Attached appendix span

≈ 9.91 page-equivalents

Before references / appendix: ≈ 41.86%References / intervening material: ≈ 13.09%Attached appendix: ≈ 45.05%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	≈ 0.99 pp.	4.5%
2. Related Work	2	≈ 0.51 pp.	2.32%
3. Geometric Analysis of Concept Erasure	3	≈ 1.01 pp.	4.59%
4. Problem Formulation	4	≈ 0.53 pp.	2.41%
5. Orthogonal Concept Erasure (OCE)	4	≈ 1.5 pp.	6.82%
6. Experiments	6	≈ 3.34 pp.	15.18%
7. Limitations.	9	≈ 0.1 pp.	0.45%
8. Conclusion	9	≈ 0.2 pp.	0.91%
Acknowledgments	10	≈ 0.05 pp.	0.23%
Impact Statement	10	≈ 0.12 pp.	0.55%
References	10	≈ 2.88 pp.	13.09%
A. Additional Theoretical Analysis	13	≈ 1.32 pp.	6%
B. Extented Related Works	14	≈ 0.28 pp.	1.27%
C. Implementation Details	14	≈ 1.75 pp.	7.95%
D. Additional Experiments	16	≈ 6.57 pp.	29.86%

All detected figure and table locations

Main text

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

Attached appendix

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

Argument and organization

1. 1 Introduction; 2 Related Work; 3 Geometric Analysis

   Uses controlled transformations to motivate preserving angular geometry during editing.

2. 4 Problem Formulation; 5 Orthogonal Concept Erasure

   Moves from additive fitting to a closed-form orthogonal objective, then relaxes erasure to the subspace level.

3. 6 Experiments; 7 Limitations; 8 Conclusion

   Tests target removal, preservation, scale, adversarial variants and architecture transfer.

Problem, claim and contribution

The core tradeoff is erasure versus retained generation quality, and the proposed restriction is motivated by a three-case geometric diagnostic rather than only leaderboard performance.

Section 3 and Figure 2, p.3 · Main paper

Argument and section order

The simpler vector-wise objective precedes the multi-concept subspace objective, with the preservation term deliberately kept vector-wise.

Sections 4-5, pp.4-6 · Main paper

Figures, captions and visual explanation

The toy panel connects three geometry operations to generated outputs; the multi-concept page aligns target/non-target examples with a table separating erasure, preservation, aggregate score and runtime.

Figure 2, p.3; Figure 3 and Table 3, p.7 · Main paper

Tables, comparisons and evidence

Target recognition rates, non-target recognition, COCO quality and runtime are reported separately; adversarial editing is explicitly marked as an added variant.

Sections 6.1-6.3 and Tables 1-4, pp.6-9 · Main paper

Appendix and reproducibility

Derivations and full target/retain sets make the closed-form objective and evaluation categories inspectable rather than leaving concept selection implicit.

Appendix A, p.13; Table 13, p.18 · Main paper

Clarity and writing conventions

The limitations acknowledge intermediate semantics instead of exact anchor matching and larger-model SVD overhead, which qualifies the language of precise erasure.

Section 7, p.9 · Main paper

What to reuse—and what not to infer

Transferable lesson: Use diagnostic interventions to motivate a constrained edit, then report removal and preservation on distinct axes and name every robustness extension.

Boundary: Preserving weight geometry is not a proof of semantic preservation or universal erasure; detector-based metrics and a few attacks define the tested scope. Runtime should distinguish editing from reusable offline preservation-prior computation.

Apply this lesson using the argument and evidence worksheet.

Inspection record and sources

Checked

2026-10-07

Reading status

Writing-focused manuscript scan + selected visuals inspected

Main text

Writing-oriented scan of the main argument through physical p.9, with focused physical-page text inspection on 1, 3, 4, 5, 6, 7, 8, 9. Framing, section progression, mechanism/proof explanation, displays, comparison conditions and conclusions were inspected. This is not exhaustive technical reading, proof verification or reproduction.

Appendix / supplement

Appendix organization and main-text pointers were checked; focused text inspection of physical pages 13, 15, 16, 18. Remaining appendix proofs and details were not exhaustively read. No separately linked supplement inspected.

Visually inspected PDF pages

3, 7

This is a writing-focused scan of the manuscript and the stated supplementary scope: framing, organization, visual explanation, evidence presentation and wording. It is not a line-by-line proof audit, reproduction of results or exhaustive scientific review.

• Published proceedings PDF, PMLR volume 306, retrieved 2026-10-07 · 22 pages