ArtifactLens: Hundreds of Labels Are Enough for Artifact Detection with VLMs
James Burgess1,
Rameen Abdal2,
Dan Stoddart2,
Sergey Tulyakov2,
Serena Yeung-Levy1,
Kuan-Chieh Jackson Wang2
1Stanford University, 2Snap Inc.
Pretrained VLMs already know how to detect artifacts in AI-generated images. With the right scaffolding, this capability can be unlocked using only a few hundred labeled examples, achieving state-of-the-art on five benchmarks while requiring orders of magnitude less data than fine-tuning.
