A method for AI image generators to find the best starting point for creating a picture. Instead of committing to one idea, it quickly explores many initial possibilities, discards the weak ones early, and only fully refines the most promising few, all while using the same amount of computing power.
Inference-Time Scaling of Diffusion Models via Progressive Seed Pruning · Rogerio Guimaraes, Pietro Perona · arXiv:2607.21591
A new AI system that can understand many different aspects of a video scene at once, like its depth, materials, and what objects are in it, even when the AI learns each aspect from completely separate training data.
Unified Video Dense Prediction from Disjoint Data · Yihong Sun, Seoung Wug Oh, Jiahui Huang et al. · arXiv:2607.21592
SANA-Video 2.0, a new AI model designed to generate high-quality videos much more efficiently. By cleverly blending two different information processing techniques, it can create longer, more detailed videos on a single powerful computer chip, making advanced video generation significantly faster and more accessible than current leading methods.
SANA-Video 2.0: Hybrid Linear Attention with Attention Residuals for Efficient Video Generation · Junsong Chen, Jincheng Yu, Yitong Li et al. · arXiv:2607.21553
An AI system that allows self-driving cars to navigate complex intersections without traffic lights by having a central "master" AI generate a compact, shared strategy for all vehicles. Individual cars then use this strategy along with their own observations to drive, leading to safer and faster travel.
Compact Latent Coordination for Autonomous Vehicles at Unsignalized Intersections · Gil Lifshits, Igal Bilik, Gilad Katz · arXiv:2607.21488
A new way for AI to learn from different types of medical scans (like X-rays and MRIs) even when the training data isn't perfectly matched, cleverly filtering out unreliable information during this process.
UnDA: Unpaired Domain Alignment for Cross-Modal Knowledge Transfer in Medical Imaging · Rafsan Jany, Shadab Tanjeed Ahmad, Ahsan Bulbul et al. · arXiv:2607.21546