ArtifactLens: Hundreds of Labels Are Enough for Artifact Detection with VLMs

TL;DR

ArtifactLens leverages pretrained vision-language models with minimal labeled data to effectively detect various image artifacts, outperforming existing methods and generalizing across multiple artifact types and detection tasks.

Contribution

The paper introduces ArtifactLens, a novel system that uses few-shot learning and text instruction optimization to detect image artifacts with minimal labeled data, surpassing prior methods.

Findings

Achieves state-of-the-art performance on five artifact benchmarks.

Requires significantly less labeled data than existing detectors.

Generalizes to diverse artifact types and related detection tasks.

Abstract

Modern image generators produce strikingly realistic images, where only artifacts like distorted hands or warped objects reveal their synthetic origin. Detecting these artifacts is essential: without detection, we cannot benchmark generators or train reward models to improve them. Current detectors fine-tune VLMs on tens of thousands of labeled images, but this is expensive to repeat whenever generators evolve or new artifact types emerge. We show that pretrained VLMs already encode the knowledge needed to detect artifacts - with the right scaffolding, this capability can be unlocked using only a few hundred labeled examples per artifact category. Our system, ArtifactLens, achieves state-of-the-art on five human artifact benchmarks (the first evaluation across multiple datasets) while requiring orders of magnitude less labeled data. The scaffolding consists of a multi-component architecture with in-context learning and text instruction optimization, with novel improvements to each. Our methods generalize to other artifact types - object morphology, animal anatomy, and entity interactions - and to the distinct task of AIGC detection.