Image-to-video (I2V) diffusion models can turn a single image into a plausible video, but that capability also makes it easy to create unauthorized content from someone else's photos. Existing defenses, known as image immunization, typically try to make the resulting video visually poor or temporally inconsistent. A NeurIPS 2026 paper instead treats the problem as one of source-content preservation: the goal is to stop the model from carrying the source image's identity into the generated video.
To do this, the proposed method, I2V-DETACH, targets the denoising process directly. I2V models propagate source-image cues into video latents as they generate frames. I2V-DETACH disrupts that propagation in two complementary ways: it suppresses source-video grounding, reducing the flow of source cues, and it strengthens intra-video interactions, pushing denoising to rely more on the video's own dynamics. The authors also initialize auxiliary video latents with source-oriented states that stay within the normal generation regime while deviating from source-preserving trajectories, which gives the optimization useful signals.
Across experiments on diverse diffusion architectures, I2V-DETACH substantially reduced source-content preservation, outperformed prior immunization methods, and exhibited strong black-box transferability. The work suggests that detaching the source grounding, rather than degrading the output, may be a more effective defense against unauthorized image-to-video generation.