Search papers, labs, and topics across Lattice.
5
0
9
0
MDN-Control, a training free framework jointly controlling target localization, occlusion geometry, and appearance initialization, is proposed, which provides consistent target localization, while depth-aware occlusion control resolves ambiguous boundaries between overlapping subjects.
A two-stage OPD-then-RL approach outperforms traditional methods by leveraging the strengths of both on-policy distillation and reinforcement learning without the interference seen in joint optimization.
Stop throwing away valuable trajectory data: a new agent can intelligently synthesize information from parallel agent rollouts, boosting performance on long-horizon tasks by up to 10%.
Image-conditioned video diffusion models can now be fine-tuned to produce more realistic motion dynamics and long-term temporal coherence via a novel reward-driven approach that avoids common pitfalls like reward hacking.
Off-the-shelf LLMs can get a 25% boost on long-context reasoning tasks simply by dynamically emphasizing relevant tokens during decoding, without any further training.
