Measuring an image without calling it style
CNNs, vision transformers, sparse features and Gram statistics help ask different questions. Their scores are evidence to interpret, not automatic artistic judgments.
Experimental · precise named controls did not qualify
Different representations, different questions.
The research architecture separates local texture-oriented CNN measurements from global ViT embeddings. A separate SAE-Gram strand investigates sparse feature activity, first-order summaries and second-order co-activation statistics. Retrieval asks whether a representation brings related examples together; intervention asks whether changing a feature produces a controlled effect.
Those tasks are not equivalent. A feature can be useful for retrieval while changing several aspects of an image when intervened on. A lower pixel error or a stronger learned similarity can also disagree with visual review.
Report the controls that did not qualify.
The inspected SAE-Gram report records no second-order retrieval advantage in the 101-image WikiArt pilot. The additional crossed cohort shows a desired separation result, while filtered Gram still trails a first-order comparison. The full real-art benchmark remains an unmet gate.
Later held-out interventions test proposed warm-palette, coarse-mark and directional-texture controls against frozen change and preservation criteria. None qualifies in the inspected report. Reference conditioning and native handoffs can execute without establishing a reliable named semantic controller.
Make measurements portable without inventing semantics.
A work-measurement attachment can carry a source identity, checkpoint and extraction configuration, sparse statistics and hashes. Its semantic and control bindings may intentionally be empty. That is a valid research artifact rather than a missing invitation to invent an interpretation.
Calibration choices, model adaptation, content leakage and donor motifs all affect interpretation. A public technical paper should show the baselines, failed criteria and uncertainty alongside attractive examples.
What the records carry
| Record | What it carries | How to interpret it |
|---|---|---|
| Feature measurement | Image, checkpoint, preprocessing and extracted vector identity | An embedding coordinate is not inherently a named art concept. |
| Calibration policy | Choices frozen before held-out evaluation | Test outcomes must not be used to quietly retune the policy. |
| Intervention report | Target change, preservation metrics and comparison conditions | Successful execution is separate from control qualification. |
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.