# BLIP-2 figure provenance

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

## generative-bridge.webp

Figure 1: Visual queries condition frozen decoder-only and encoder-decoder models.

Figure 3, published PDF page 4, crop x=322 y=272 width=1848 height=552 at 300 dpi.

Source: Junnan Li, Dongxu Li, Silvio Savarese, Steven Hoi, BLIP-2 report. https://proceedings.mlr.press/v202/li23q.html

License: [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/). This report artwork is not covered by the repository MIT blanket. Changes: cropped to the named figure or panel, excluding paper prose/caption; data, axes, legend and diagram content unchanged. Rendered at 300 dpi and encoded as lossless WebP.

Permission evidence: https://arxiv.org/abs/2301.12597v3

## qformer-objectives.webp

Figure 2: Q-Former architecture and the three query-text attention masks.

Figure 2, published PDF page 3, crop x=216 y=282 width=2057 height=421 at 300 dpi.

Source: Junnan Li, Dongxu Li, Silvio Savarese, Steven Hoi, BLIP-2 report. https://proceedings.mlr.press/v202/li23q.html

License: [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/). This report artwork is not covered by the repository MIT blanket. Changes: cropped to the named figure or panel, excluding paper prose/caption; data, axes, legend and diagram content unchanged. Rendered at 300 dpi and encoded as lossless WebP.

Permission evidence: https://arxiv.org/abs/2301.12597v3

Selected arXiv v3 metadata explicitly links CC BY 4.0. Published PMLR paper is also covered by the venue publication agreement clauses 2 and 3. Source figure permissions are separate from LAVIS, OPT and model artifact terms. No figure-specific contrary rights notice found in the inspected report.
