Search papers, labs, and topics across Lattice.
6
0
9
Achieving a 359-degree median span in 3D scene reconstruction from text prompts could redefine the standards for text-to-3D generation.
Camera-prompted T2V models often fail to align visual motion with static scene coherence, revealing hidden inconsistencies in their outputs.
Self-distillation enables visual world models to solve tasks without costly video supervision, achieving superior performance in both simulated and robotic environments.
Soft attention-discounted reranking can boost masked diffusion model performance by over 10 percentage points while maintaining efficient decoding.
Squeeze 2x more speed from your conditional flow matching models by optimizing data-noise coupling across minibatches.
Object-centric image tokenization is now possible by modeling the sequential refinement process of human communication, leading to better compositional generalization and relational reasoning.
