# Qwen2-VL figure provenance

Checked October 10, 2026. Full archival PDFs stay outside the public repository.

## visual-interface.svg

Page Figure 2: Patch-grid compression and decoder sequence accounting.

Selected processor, PatchMerger and worked image counts; Independent token-accounting diagram; the report framework is separately reproduced under the revision 4 assessment below.

Original Latent Light SVG, independently authored from the selected model configuration and code described in the corresponding study. MIT licensed with this repository. No source artwork was copied or traced.

Selected report license links arXiv nonexclusive distribution, not a public reuse grant. Official release README at d90d8e31 links figure assets hosted externally, rather than distributing those diagrams as repository files. Blog repository lacks an overall content license (only theme LICENSE). The SVG is independently authored, not traced. Revision 4 separately reproduces the original report framework under the assessment below.

## mrope-coordinates.svg

- Original, independently authored vector diagram; not traced or copied report artwork. Latent Light, 2026. MIT under the repository license.
- Source basis: Qwen2-VL arXiv 2409.12191v2, section 2.1, plus Transformers 4.57.1 get_rope_index at revision 8cb5963cc22174954e7dca2c0a3320b7dc2f4edc. Toy 2 × 3 merged grid with first visual coordinate five; exact coordinate recurrence, not a processor benchmark.
- Purpose: Separate geometric coordinate values, physical sequence positions and ordinal video tube time.
- Static light canvas; accessible title, description, page alt text and prose equivalent. Revision 3 browser review completed; revision 4 review is recorded in the research state.

## Revision 4: original report figure excerpts

These specific excerpts support substantive model analysis on the public site
and in its public source repository. Original copyright remains with the
respective rights holders; no MIT or other new artwork license is granted.
A statutory exception assessment is not permission from the authors or a legal
assurance for another use or jurisdiction. Reassess redistribution, adaptation
or commercial use separately.

### report-architecture.png

- Source: Peng Wang et al.. [Qwen2-VL: Enhancing Vision-Language Model's Perception of the World at Any Resolution](https://arxiv.org/pdf/2409.12191v2), original Figure 2, PDF p. 4; exact version is in the URL.
- Purpose: Analyze dynamic resolution and how unlike image/video inputs become unequal visual token groups in one text decoder.
- Changes: high-resolution 300 dpi figure-region crop; paper prose and caption omitted. All figure panels, labels and arrows retained, with no redraw or AI alteration.
- Crop pixels: x 287, y 291, width 1980, height 1267.
- Reuse basis: case-specific Canadian fair-dealing assessment for research and criticism/review, not a permissive-license claim.
- Third-party scope: Embedded blog, landscape and sailing-video thumbnails have no separately established grant. They are retained at diagram-input scale for analysis of the original processing paths, not as reusable photography.
- Source PDF SHA-256: 30bb0f6c9babf910295d6b7cd34b6f54824236f68782664c3c93fbf4803275d0
- Asset SHA-256: bb30803b667e0f7677630db54a2ee9321993395c7421406bd1d1f5cd326fcb7d

### Assessment for this selection

- Purpose: explain and critically inspect the published model mechanisms beside the figures, with author and source attribution. The figures are evidence, not decoration.
- Character: publicly accessible web and GitHub copies are broader distribution than private study and weigh against reuse; they must be considered explicitly. No complete paper, separate photographic asset or standalone gallery is distributed.
- Amount: 1 complete figure from the 52-page report. Each figure may itself be a complete protected work; report length alone is not the justification. All panels are needed to inspect the depicted relationships, while surrounding prose and unrelated figures are omitted.
- Alternatives: links and independent schematics remain useful, but are not equivalent for inspecting the authors' exact grouping, arrows and presentation. Original figure reproduction is paired with a focused explanatory passage, not used to avoid writing analysis.
- Nature: an already publicly released technical report, not confidential or unpublished work. Public access does not itself grant reuse.
- Effect: the excerpts link readers to the original edition and do not replace its full argument, experiments or report. No measured market-effect study was performed; a competing use, rightsholder concern or monetization would require reassessment.
- Decision: include these bounded excerpts for the documented analysis. This editorial assessment has remaining jurisdictional and rights uncertainty; it is not a judicial determination or a grant to downstream users.

Legal sources: [Copyright Act §29](https://laws-lois.justice.gc.ca/eng/acts/c-42/Section-29.html), [§29.1](https://laws-lois.justice.gc.ca/eng/acts/c-42/Section-29.1.html), and [CCH Canadian, 2004 SCC 13](https://decisions.scc-csc.ca/scc-csc/scc-csc/en/item/2125/index.do), paras. 48–60. Assessed October 10, 2026.
