Caesar can illustrate a finished report using images found on its cited pages. The pipeline revisits those sources, filters and scores candidate images, then uses selected references to help write a generation prompt.
In a run about biomimetic architecture, four references contributed recognizable visual details. One incorrect description of stone columns as wood also reached the final image.
From cited pages to four references
| Stage | Count |
|---|---|
| Cited pages revisited | 20 |
| Candidates after scraping filters | 343 |
| Candidates scored by the vision model | 60 |
| References retained | 4 |
(Reported counts and plotting code.)
The scraper collects image candidates from cited pages, removes duplicate URLs
and filters known boilerplate path segments such as logo, icon, favicon,
avatar and pixel. The 343 candidates are what remain after filtering.
There is no size filter, so decorative headers, author photographs and other
irrelevant material can survive.
The pipeline sends at most 60 candidates to the vision model. This run selected 60 without relevance ranking; 283 candidates were never visually assessed. Each selected candidate received a score from 0 to 10 for suitability as a reference for the report’s opening. Candidates below 4.0 were discarded; the rest were ranked, with at most four retained and at most two from one domain.
A later implementation first ranks candidates using text from alt attributes and URL keywords, selects the top 50, and samples 10 from the remaining pool. The random component allows images with poor metadata to enter consideration. Both Wikimedia references retained in the earlier run had empty alt text.

The retained references scored 8.0, 8.0, 8.0 and 7.0. Two were Wikimedia Commons files embedded in a cited Wikipedia article. Two came from a video-intercom company’s blog, with alt text describing tree-inspired architecture and a biomimetic city skyline.
A vision model described each reference. A language model then combined those descriptions with the report to write an image prompt, specifying an upward interior view, branching pale-wood ribs, faceted canopy panels and circular skylights.
Tracing visible details

The image contains lattice towers and planted high-rises resembling details in the selected references. An annotated deep-sea sponge photograph offers a more specific trace. Its caption described red arrows and white markings as giving the image a documentary appearance; the prompt requested a documentary inset, and the output contains a labeled technical panel.
The visual matches below were inferred from the recorded descriptions, prompt and visible output. The pipeline does not log which reference controls each output region; the numbered provenance overlay was matched by eye. The architecture run’s scores, detailed descriptions and prompt come from unpublished run records.

The vision model misread the Sagrada Família reference. Its branching columns are stone, but the vision description called them pale wood. That material entered the prompt and appears in the output. Reusing sources can transmit misreadings along with useful details; visual resemblance alone cannot validate the result.
Other reports and fallback behaviour
The same pipeline processed four earlier reports on a later build. Their outputs vary substantially in style, but the examples do not measure image quality or isolate the contribution of references.

The prompt-writing model normally chooses among twelve visual idioms supplied by the template, including etching, lithograph, blueprint, cyanotype and specimen plate. It matches an idiom to the subject and borrows palette or materials from reference descriptions. A separate diagram mode, triggered by patterns in the report text, forces a flat technical presentation and excludes reference images.
The creativity-scoring report produced 171 candidates, with a best score of 4.0. That falls below the 7.0 best-reference gate required to use images. The image was generated from text alone. Diagram mode had already disabled references, so the score gate did not determine that choice. This example cannot establish that the gate improved the result.
Selection settings and their limits
The 4.0 individual-candidate cutoff and 7.0 best-reference gate have different roles. Below the latter, the captioning pass is skipped and the prompt is written from the report alone. Above it, selected references can supply both visual input and descriptions.
The 7.0 gate is intended to avoid distracting references, but the examples do not establish that this threshold is optimal.
A configuration comment reports another selection failure: one vision scorer clustered scores near the rubric’s middle, leaving little signal for reranking. This is an implementation observation, not a controlled model comparison.
The pipeline reuses material gathered during research, much as saved exploration snapshots enable later analysis. Its examples show source-specific visual details and identifiable error propagation. A matched comparison with text-only generation would be needed to establish whether the additional scraping, scoring and captioning improve illustrations overall.