Images as Tables: In-Context Learning with TabPFN for Low-Data Detection of AI-Generated Images

Jan Philip Walter, Shashank Agnihotri, Margret Keuper · 2026

Read Images as Tables: In-Context Learning with TabPFN for Low-Data Detection of AI-Generated Images by Jan Philip Walter, Shashank Agnihotri, Margret Keuper (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2606-00872.txt.

Description

Walter, Agnihotri, and Keuper turn fake-image detection into a table task in this 2026 paper. Detectors trained on one image generator often fail on the next one. Labeled examples of the new fakes are few. Retraining for each shift is slow and costly.

Their recipe reads each image with a frozen DINOv3 backbone. A PCA step squeezes its feature vector into a 500-number row. TabPFN, a tabular foundation model, then classifies rows by in-context learning, with no task-specific training. On the GenImage set, a top detector named LATTE wins when many labels from all generators exist, by 7.4 percent. But in the low-data regime that matters in practice, the table method wins by up to 8.2 percent. It also transfers better from one generator to another. Adaptation becomes a small labeled set, not a retraining run.

Catalog record

CREATORJan Philip Walter, Shashank Agnihotri, Margret Keuper
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE32 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2606.00872

Browse the full library · About the Tower of Babel