The corpus behind the questions
How artwork images, metadata and art-historical text enter different parts of the research—and why the identity of every input matters.
Documented methods · incomplete full-corpus evaluation
A corpus is more than a folder of pictures.
The project uses several kinds of evidence. WikiArt-labeled image cohorts support visual-representation and retrieval experiments. Saved artwork descriptions and museum sources support cited concept discovery. Individual painting studies use explicitly admitted reference images. These collections have different purposes and cannot be treated as one interchangeable training set.
An artwork, a reproduction of it, a crop and a model-ready image are different identities. Artist and movement labels describe a dataset record; they do not by themselves establish the correctness of an attribution, the independence of a sample or permission to redistribute the image.
WikiArt: cohorts, not a blanket training claim.
The inspected SAE-Gram research documents a completed 101-image WikiArt pilot and a separate 240-image crossed cohort. They examine whether visual representations retrieve works with related labels and separate style signals from content. Completion of those studies is distinct from successful full-dataset evaluation.
A larger all-published-row protocol is recorded separately, with image admission and complete evaluation still unmet in the inspected evidence. Its expected row count is a protocol target, not an observed number of images processed. We do not describe this as a model trained on all of WikiArt.
Preserve the route from source to measurement.
The research records source-byte, decoded-image and prepared-image identities, preprocessing choices and evaluation roles. Duplicate exclusions and calibration overlap matter because a query that matches its own reproduction is not useful evidence of generalization.
Text records keep source locators and retained passage identities. Multiple excerpts from one provider are not automatically independent corroboration. Public publication requires a separate review of image rights and source permissions; the technical notes do not distribute the research image corpus.
What the records carry
| Record | What it carries | How to interpret it |
|---|---|---|
| Image manifest | Source identity, labels, preparation and split membership | A label is not a validated artistic trait. |
| Source record | Text identity, locator, retained passage and source grouping | A retained excerpt is not the complete source page. |
| Admission receipt | Checks against the exact files used by a study | An expected manifest does not prove the images were admitted. |
Methods summary from the inspected research snapshot. Findings described here belong to their stated experiments; this note is not a fresh replication or a peer-reviewed publication.