Writing study: Repurposing Geometric Foundation Models for Multi-view Diffusion
← ECCV 2026 oral writing corpus
Wooseok Jang, Seonghu Jeon, Jisang Han, Jinhyeok Choi, Minkyung Kwon, Seungryong Kim, Saining Xie, Sainan Liu · Official ECCV 2026 oral record
Figure, table and section ledger
Editorial contribution type: Architecture & representation (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
- 19 physical pages
- Before references / appendix
- ≈ 15.15 page-equivalents
- Main figures / tables
- 7 / 9
- Attached appendix figures / tables
- 0 / 0
- Attached appendix span
- No boundary detected; not assumed zero
Extraction notes: No attached appendix boundary detected. Zero means no recognized caption, not independently verified absence.
Source-bound correction: All top-level main headings and the reference boundary checked against source PDF text; source-bound positions use physical single-column y/page-height, not a two-column reading-order surrogate. Running page-number/author titles, affiliations, numbered contribution bullets, footnotes and bibliography continuations are excluded. This linked manuscript contains no appended supplementary section; separate supplement links were not scanned. Missing split/side-wrapped caption labels restored after rendered-page inspection.
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 | 2 | ≈ 1.05 pp. | 5.53% |
| 2 Related Work | 3 | ≈ 0.53 pp. | 2.79% |
| 3 Preliminaries | 3 | ≈ 0.88 pp. | 4.63% |
| 4 Method | 4 | ≈ 2.98 pp. | 15.68% |
| 5 Experiments | 7 | ≈ 7.09 pp. | 37.32% |
| 6 Conclusion | 14 | ≈ 0.55 pp. | 2.89% |
| Acknowledgements | 15 | ≈ 0.87 pp. | 4.58% |
| References | 16 | ≈ 3.85 pp. | 20.26% |
All detected figure and table locations
Main text
Attached appendix
No recognized caption labels.
Argument and organization
- 1 Introduction; 2 Related Work; 3 Preliminaries
Asks which latent space serves novel-view synthesis, separating its geometric consistency needs from single-image generative representations.
- 4 Method: reconstruction validation, boundary selection and cascade
Validates RGB reconstructability before choosing a feature synthesis boundary, then tests cascaded generation; later diagnostics revisit that boundary from two complementary perspectives.
- 5 Experiments and analyses; 6 Conclusion
Comparisons separate in-domain and out-of-domain data, from-scratch and pretrained initializations, and 2D image from 3D consistency metrics.
Problem, claim and contribution
Asks which latent space serves novel-view synthesis, separating its geometric consistency needs from single-image generative representations.
Abstract and Section 1, pp.1–2 · Main paper
Argument and section order
Validates RGB reconstructability before choosing a feature synthesis boundary, then tests cascaded generation; later diagnostics revisit that boundary from two complementary perspectives.
Sections 4.2–4.4, pp.5–7; Section 5.5, p.12 · Main paper
Figures, captions and visual explanation
The overview distinguishes explicitly generated feature levels from levels propagated by a frozen geometric model; reconstruction crops sit beside layer-specific geometry and image-fidelity tables.
Figure 2, p.5; Tables 5–6 and Figure 5, p.12 · Main paper
Tables, comparisons and evidence
Comparisons separate in-domain and out-of-domain data, from-scratch and pretrained initializations, and 2D image from 3D consistency metrics.
Table 3, p.8; Section 5.2, pp.9–10 · Main paper
Appendix and reproducibility
The manuscript points to named appendix sections for training, boundary evaluation, an alternative geometric backbone and qualitative results. Those sections were not separately retrieved.
Sections 4–5, pp.4,7–8,11 · Main paper
Clarity and writing conventions
Short subsection headings are phrased as design decisions to validate, and tables make level-specific trade-offs explicit instead of describing deeper features as universally better.
Sections 4.2–4.4; Section 5.5, pp.5–7,12 · Main paper
What to reuse—and what not to infer
Transferable lesson: Before proposing a generative latent space, demonstrate what it can reconstruct and what it preserves; connect the chosen boundary to both diagnostics.
Boundary: The source's convergence, reconstruction and novel-view comparisons use different protocols. Better correspondence at one feature level need not imply better RGB reconstruction.
Apply this lesson using the argument and evidence worksheet.
Inspection record and sources
- Checked
- 2026-10-02
- Reading status
- Writing-focused manuscript scan + selected visuals inspected
- Main text
- Writing-oriented scan across the manuscript main argument, physical PDF pp.1–14: framing and section roadmaps, method/argument explanations, figure and table captions, comparison protocols, conclusions and limitations. Selected full-page rasters were inspected. This is not exhaustive line-by-line technical reading, proof verification or experimental reproduction.
- Appendix / supplement
- Main-manuscript pointers to appendix/supplement and the reference boundary were checked. Separate supplementary PDFs were not retrieved or inspected; their contents are not included in inspected-page counts.
- Visually inspected PDF pages
- 5, 12
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.