Valerant: An Automatic NaVigAbLE Game Map GeneratoR via Action-CoNditioned World Model ExploraTion

Yiran Qiao, Feng Wang, Jing Ma

[Paper] [Project Page]

• VALERANT builds a navigable 3D game map from a single image by rolling out candidate actions through an action-conditioned world model, reconstructing each rollout with SLAM, and committing the step that scores best.

Structured 3D Latents Are Surprisingly Powerful: Unleashing Generalizable Style with 2D Diffusion

Yiran Qiao, Yiren Lu, Yunlai Zhou, Disheng Liu, Linlin Hou, Rui Yang, Yu Yin, Jing Ma

[Paper] [Project Page]

• We introduce DiLAST: 2D Diffusion-based Latent Awakening for 3D Style Transfer. It guides a 3D generator’s structured latents with a pretrained 2D diffusion teacher, transferring styles well outside the generator’s training distribution and plugging into existing backbones unchanged.

AdvSplat: Adversarial Attacks on Feed-Forward Gaussian Splatting Models

Yiran Qiao, Yiren Lu, Yunlai Zhou, Rui Yang, Linlin Hou, Yu Yin, Jing Ma

[Paper] [Project Page]

• AdvSplat gives the first systematic account of adversarial attacks on feed-forward 3DGS, with two query-efficient black-box algorithms that parameterize imperceptible pixel perturbations in the frequency domain and disrupt reconstruction without access to model internals.

DefenseSplat: Enhancing the Robustness of 3D Gaussian Splatting via Frequency-Aware Filtering

Yiran Qiao, Yiren Lu, Yunlai Zhou, Rui Yang, Linlin Hou, Yu Yin, Jing Ma

[Paper] [Code]

• DefenseSplat filters poisoned input views in the wavelet domain and regularizes Gaussian scale during training, restoring 3DGS fidelity under attack without adversarial training or clean ground truth.

Certified Causal Defense with Generalizable Robustness

Yiran Qiao, Yu Yin, Chen Chen, Jing Ma

[Paper] [Code]

• GLEAN learns causally invariant features behind an L-Lipschitz encoder, certifying a robustness radius that holds across shifted domains.