Writing study: Improving Diffusion Models for Class-imbalanced Training Data via Capacity Manipulation

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Feng Hong, Jiangchao Yao, Yifei Shen, Dongsheng Li, Ya Zhang, Yanfeng Wang · 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) · Theory & formal foundations.

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

40 physical pages

Before references / appendix

≈ 10.2 page-equivalents

Main figures / tables

4 / 7

Attached appendix figures / tables

12 / 15

Attached appendix span

≈ 23.88 page-equivalents

Before references / appendix: ≈ 25.5%References / intervening material: ≈ 14.8%Attached appendix: ≈ 59.7%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	≈ 0.98 pp.	2.45%
2 ON THE LIMITATIONS OF DIFFUSION MODELS ON IMBALANCED DATA	2	≈ 1.25 pp.	3.13%
3 METHOD: CAPACITY MANIPULATION	3	≈ 2.62 pp.	6.55%
4 EXPERIMENTS	6	≈ 4.29 pp.	10.72%
5 CONCLUSION	10	≈ 0.33 pp.	0.83%
ACKNOWLEDGEMENTS	11	≈ 0.09 pp.	0.22%
REFERENCES	11	≈ 5.93 pp.	14.82%
A PPENDIX : I MPROVING D IFFUSION M ODELS FOR	17	≈ 2.48 pp.	6.2%
B RELATED WORK	19	≈ 0.93 pp.	2.33%
C ALGORITHM PSEUDOCODE	20	≈ 0.58 pp.	1.45%
D BRIEF INTRODUCTION TO THE OBJECTIVES OF CBDM AND OC	21	≈ 0.51 pp.	1.28%
E MORE DISCUSSION	21	≈ 1.22 pp.	3.05%
F THEORETICAL SUPPLEMENT	22	≈ 3.67 pp.	9.18%
G EXPERIMENTAL SUPPLEMENT	26	≈ 13.17 pp.	32.92%
H SOCIAL IMPACT	39	≈ 0.42 pp.	1.05%
I LIMITATIONS, DISCUSSIONS, AND FUTURE WORK	40	≈ 0.24 pp.	0.6%
J USE OF LARGE LANGUAGE MODELS (LLMS)	40	≈ 0.66 pp.	1.65%

All detected figure and table locations

Main text

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

Attached appendix

• Table E.1 · p. 22
• Table E.2 · p. 22
• Figure G.1 · p. 26
• Figure G.2 · p. 27
• Table G.1 · p. 27
• Table G.2 · p. 28
• Table G.3 · p. 28
• Table G.5 · p. 28
• Table G.6 · p. 28
• Table G.4 · p. 29
• Table G.7 · p. 29
• Figure G.3 · p. 29
• Table G.8 · p. 30
• Figure G.4 · p. 30
• Figure G.5 · p. 31
• Figure G.6 · p. 32
• Figure G.7 · p. 32
• Figure G.8 · p. 33
• Figure G.9 · p. 34
• Figure G.10 · p. 35
• Figure G.11 · p. 36
• Figure G.12 · p. 37
• Table G.9 · p. 38
• Table G.10 · p. 38
• Table G.11 · p. 38
• Table G.12 · p. 38
• Table G.13 · p. 39

Argument and organization

1. Motivation and theoretical insights

   Connects pruning sensitivity to a restricted gradient-dominance model.

2. Capacity Manipulation

   Separates parameter decomposition from the class-dependent allocation loss.

3. Experiments and further analysis

   Tests total quality, minority quality and component removal separately.

Problem, claim and contribution

The opening diagnostic is specific: similar raw diffusion losses coexist with minority sensitivity to removing small-magnitude parameters. This motivates a capacity-allocation hypothesis without making loss imbalance the sole explanation.

§2.2 and Figure 1, p.3; Appendix G.2, pp.26–27 · Main paper

Argument and section order

The method distinguishes reserving low-rank parameters from assigning them minority expertise. Comparisons against decomposition alone, fine-tuned LoRA and group experts directly address whether extra parameters explain the improvement.

§3, pp.4–6; Table 1; Appendix E.3–E.4, p.22 · Main paper

Figures, captions and visual explanation

Figure 1 pairs generated examples with a pruning-response curve; Table 3 then separates many, medium and few classes instead of hiding the minority effect inside overall FID. Captions state imbalance ratios and group definitions.

Figure 1, p.3; Tables 2–3, p.7 · Main paper

Tables, comparisons and evidence

The evaluation explicitly uses balanced real reference images, 50,000 generated images overall and 20,000 per group. Reporting this pipeline prevents comparisons against unrelated balanced-training FID values from appearing controlled.

§4.1–4.2, pp.6–7 · Main paper

Appendix and reproducibility

The supplement supplies Gaussian-gradient assumptions, an idealized low-rank update model, pseudocode, standard deviations, human preferences and downstream augmentation. Its scarcity limitation usefully distinguishes imbalance ratios from having only a handful of minority examples.

Appendices C, F, G.3–G.5 and I, pp.20–30,40 · Main paper

Clarity and writing conventions

Broad phrases such as universal framework should be tied to the tested objectives and architectures. The discriminative extension is preliminary, and its group-wise table includes trade-offs rather than uniform improvement for every group.

§4.3, pp.9–10; Table G.13, p.39 · Main paper

What to reuse—and what not to infer

Transferable lesson: Use an explanatory diagnostic, a mechanism-specific control and a disaggregated outcome table to make a minority-performance argument testable.

Boundary: Parameter magnitude is a proxy, not a causal map of stored knowledge. Theorems depend on simplified Gaussian gradients and update normalization; they do not establish the actual loss's universal guarantees. Zero overhead concerns merged inference, not training. CM is joint capacity allocation, not teacher–student distillation.

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.1–10, §§1–5; references pp.11–16

Appendix / supplement

pp.17–40, appendices A–J including assumptions, proofs, pseudocode, supplementary experiments and limitations

Visually inspected PDF pages

3, 7

• official ICLR 2026 proceedings PDF · 40 pages