Writing study: VAEVQ: Enhancing Discrete Visual Tokenization Through Variational Modeling
← AAAI 2026 oral writing corpus
Sicheng Yang, Xing Hu, Qiang Wu, Dawei Yang · Official AAAI 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
- 9 physical pages
- Before references / appendix
- ≈ 7.04 page-equivalents
- Main figures / tables
- 7 / 3
- 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: Checked 2026-10-25 against the manuscript scan and physical-page Poppler boxes. Restore unnumbered main headings; VLQ/RCS/DCR, Visual Generation Performance and Ablation Study are subordinate headings, not new main sections. Captions match Figures 1–7 and Tables 1–3. Physical pp.3,6,7 were rendered and visually inspected. The publisher PDF ends with references and has no attached appendix.
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.
All detected figure and table locations
Main text
Attached appendix
No recognized caption labels.
Argument and organization
- Introduction; Related Work
Identifies irregular latent geometry, pre/post-quantization mismatch and poor codebook coverage as separate limitations, then relates deterministic tokenizers to the proposed variational alternative.
- Methodology
Maps the three limitations to variational latent quantization, uncertainty-aware representation coherence and codebook distribution regularization; the shared decoder and objective equations make the pathways explicit.
- Experiments; Conclusion
Tests reconstruction and codebook utilization before downstream generation, then varies component combinations and codebook size and acknowledges the limitation of a fixed-size codebook.
Problem, claim and contribution
The introduction distinguishes three bottlenecks instead of using codebook collapse as a single catch-all explanation. It maps latent geometry, quantization-boundary alignment and global codeword distribution to VLQ, RCS and DCR respectively. This creates a claim-to-module contract that the final ablation table can test with the same names.
Introduction and Figure 1, physical pp.1-2; Methodology opening, pp.2-3; Table 3, p.7 · Main paper
Argument and section order
The method starts with the changed representation path: predict a Gaussian posterior, sample a continuous latent, quantize it and decode both through shared weights. Only then does it introduce local uncertainty-weighted alignment and a global codebook-distribution penalty. Experiments retain that separation by measuring reconstruction, codebook usage and downstream generation as different outcomes.
Figure 2 and Equations 1-4, p.3; RCS/DCR, pp.4-5; reconstruction/generation results, pp.5-7 · Main paper
Figures, captions and visual explanation
Figure 2 is a full-width architecture with a stochastic latent path, a quantized path and a shared decoder; two insets explain local RCS and global DCR instead of adding every mathematical term to the main flow. Figure 6 uses matched image columns and red regions of interest for natural images and MRI slices. Figure 7 separates utilization above reconstruction/generation curves, but its lower dual-axis panel requires careful axis labeling rather than direct vertical comparison between rFID and gFID.
Figure 2, physical p.3; Figure 6, p.6; Figure 7, p.7; rendered pages inspected · Main paper
Tables, comparisons and evidence
Table 1 groups reconstruction PSNR, SSIM and rFID by dataset and prints metric directions. Table 2 instead compares generation FID after substituting the tokenizer into two generative backbones. Table 3 exposes component switches rather than only reporting the full model. Figure 7 shows diminishing returns as codebook size increases, supporting the stated choice instead of suggesting unlimited scaling gains.
Training setup, p.5; Table 1 and Figure 5, p.6; Tables 2-3 and Figure 7, p.7 · Main paper
Appendix and reproducibility
The main manuscript contains the stochastic reparameterization, quantization rule, dual-path reconstruction and both regularizers, along with dataset split, image size, codebook size and hardware. No appendix is attached to this PDF. A reproducibility package should make slice extraction, subject-level splitting, model revisions and the generation evaluation configuration directly locatable; repository contents were not inspected here.
Equations and methodology, pp.3-5; Datasets and Implementation Details, p.5; references, pp.8-9 · Main paper
Clarity and writing conventions
The VLQ/RCS/DCR names remain consistent from the motivating gaps to the architecture and ablation columns. Maintain equally precise numerical language: SSIM from 0.66 to 0.68 is an absolute increase of 0.02, not automatically a two-percent relative gain. BraTS24 MRI-slice results are a second imaging-domain test, not evidence about whole-slide pathology or clinical utility.
Methodology, pp.3-5; Table 1 and quantitative discussion, p.6; component table and limitation, p.7 · Main paper
What to reuse—and what not to infer
Transferable lesson: For a representation architecture with several objectives, keep the main data path visually dominant and place local/global penalties in distinct insets. Reuse module names in the component table, distinguish reconstruction from downstream generation, and report the parameter regime where scaling benefits flatten.
Boundary: A conceptual Gaussian diagram is not empirical proof that learned codewords follow that distribution. Reconstruction and generation FID are different evaluations; the comparison tables do not report uncertainty. The MRI experiments are not WSI or clinical validation, and near-complete codebook usage alone does not establish better semantic representations.
Apply this lesson using the argument and evidence worksheet.
Inspection record and sources
- Checked
- 2026-10-25
- Reading status
- Writing-focused manuscript scan + selected visuals inspected
- Main text
- Writing-focused scan of physical PDF pp.1-7 covering the three-part motivation, variational tokenizer, shared decoding pathways, local/global alignment objectives, reconstruction and generation evaluations, component/codebook-size ablations and limitation. Physical pp.3, 6 and 7 were rendered and visually inspected; the reference boundary on p.8 was checked. The mathematical modeling assumptions were read for their argumentative role, not independently proved.
- Appendix / supplement
- The inspected nine-page proceedings PDF has no attached appendix; references occupy pp.8-9. No main-text appendix pointer was identified. The article's separate video/poster/slides asset and linked code repository were not inspected, so availability of additional settings or experiments is not inferred from these links.
- Visually inspected PDF pages
- 3, 6, 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.