Tongliang Liu

Existing safety guardrails for text-to-image models can backfire, inadvertently amplifying other types of harm, but this new method adaptively steers generation to resolve these conflicts and reduce overall harmful content.

Image-to-video models can be jailbroken by hiding malicious instructions in seemingly harmless reference images, achieving an 83.5% attack success rate on commercial systems.

By explicitly disentangling target features with MLLM guidance, MeGU achieves superior unlearning performance without sacrificing model utility, outperforming existing methods that struggle with the inherent entanglement of semantic concepts in model representations.

Forget finicky time steps: i-PhysGaussian uses implicit integration with 3D Gaussian Splatting to simulate physics with 20x greater stability.