Code Paper PIXAR studies image tampering from masks to pixels and semantics for vision-language models.