Anti-Prompt: Image Protection against Text-Guided Image-to-Video Generation

2026-07-08 · Source: cs.CV updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cybersecurity & Data Privacy · Depth: Expert, extended

Summary

Anti-Prompt is an image protection method designed to counter unauthorized text-guided Image-to-Video (I2V) generation, addressing copyright and privacy risks. It injects imperceptible perturbations into an image, inducing visible inconsistencies and structural failures in generated videos. The method exploits I2V models' reliance on textual guidance by attenuating text-conditioned interactions during denoising while strengthening visual-only pathways. Anti-Prompt was evaluated on CogVideoX-5B (full-attention) and LTX-Video-2B (cross-attention) architectures, demonstrating strong protection performance, improved efficiency, and cross-model transferability. A Video-LLM-assisted evaluation protocol, assessing Subject Preservation, Structural Consistency, Dynamic Consistency, and Artifact Suppression, was introduced to systematically analyze generation artifacts.

Key takeaway

For AI Security Engineers or content creators concerned about unauthorized video generation from shared images, Anti-Prompt offers a robust defense. You should consider implementing this method to inject imperceptible perturbations into your images, thereby inducing visible inconsistencies and structural failures in text-guided I2V outputs. This proactive approach helps mitigate privacy and copyright risks by making generated videos unsuitable for convincing misuse, even against unseen prompts and different I2V models.

Key insights

Anti-Prompt protects images from I2V misuse by subtly perturbing them to weaken text guidance and amplify visual-only pathways.

Principles

• I2V models degrade significantly without text guidance.
• Text guidance stabilizes I2V generation broadly.
• Protection can target architectural text pathways.

Method

Anti-Prompt optimizes imperceptible perturbations by suppressing text-dependent attention interactions and reinforcing visual-only pathways, complemented by an encoder attack to distort image conditioning.

In practice

• Apply perturbations to images before sharing online.
• Use Video-LLMs for nuanced failure detection.
• Consider Anti-Prompt for cross-model I2V protection.

Topics

• Image Protection
• Image-to-Video Generation
• Diffusion Models
• Adversarial Perturbations
• Video-LLM Evaluation
• CogVideoX
• LTX-Video

Best for: Research Scientist, CTO, VP of Engineering/Data, AI Scientist, Machine Learning Engineer, AI Security Engineer

Related on AIssential

• Anti-Prompt: Image Protection against Text-Guided Image-to-Video Generation — Anti-Prompt protects images from text-guided I2V generation by exploiting model reliance on textual guidance.
• VPA-Guard: Defending and Benchmarking Image-to-Video Generation Against Visual Prompt Attacks — Visual prompt attacks pose a severe threat to Image-to-Video models, necessitating specialized benchmarks like VVA-Bench and defenses like VPA-Guard.
• AEGIS: A Mechanism-Guided Defense against Visual Synonym Jailbreaks in Text-to-Image Models — Unsafe visual semantics in T2I models emerge via sparse "semantic-injecting attention heads" that can be dynamically steered.
• Efficient, Robust, and Anti-Collusion Fingerprinting of Image Diffusion Models — A novel T2I model fingerprinting method uses a Personalized Normalization Module and anti-collusion transformations to ensure robustness against malicious model combinations.
• Florida International University researchers reveal how altered images can bypass AI safeguards — Microscopic image perturbations can "jailbreak" small-language AI models, bypassing safety mechanisms and generating harmful outputs.
• Prompt-inject ChatGPT with any web page — Chatbots inherently struggle to differentiate instructions from data, making prompt injection an unfixable security vulnerability.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.