VETO cloak blocks AI image editing by disrupting FLUX.2's joint-attention
New VETO cloak breaks modern AI editors like FLUX.2 with subtle pixel changes.
Researchers from TU Darmstadt (Jonas Grebe, Hossein Shakibania, Tobias Braun, Marcus Rohrbach, Anna Rohrbach) introduced VETO, a novel anti-edit cloak designed to protect images from frontier AI editing models. Modern editors like FLUX.2 have evolved beyond simple localized edits—they can extract objects and identities and recontextualize them into entirely new scenes. This capability stems from joint-attention blocks that allow prompt and generation tokens to attend directly to reference-image tokens, blurring the line between editing and text-to-image synthesis. Existing defenses, which target the semantic bottleneck in legacy diffusion pipelines, often fail against these newer architectures.
VETO works by subtly perturbing images to disrupt the joint-attention mechanism, preventing the model from effectively reading the source image. To evaluate such defenses comprehensively, the team also created VetoBench, a benchmark that tests both conventional localized edits and broader contextual shifts like recontextualization. Across two contemporary editing models and three benchmarks, VETO consistently outperformed existing defenses while offering a stronger protection-fidelity trade-off, meaning it protects images without visibly degrading their quality.
- VETO disrupts joint-attention blocks in modern editors like FLUX.2, bypassing defenses aimed at legacy diffusion pipelines.
- Introduces VetoBench, a benchmark that evaluates both localized edits and full recontextualizations.
- Outperforms existing defenses across two editing models and three benchmarks, with a better protection-fidelity trade-off.
Why It Matters
VETO gives photographers and artists a practical tool to stop AI from stealing or recontextualizing their work.