Spyros Gidaris

End-to-end learning of temporally predictable latent spaces for world models unlocks superior future scene understanding without the need for separate representation and prediction training stages.

Synthetic scenes with automatically grounded object identities unlock substantial real-world perceptual gains in MLLMs via on-policy self-distillation.

Image diffusion models get a boost by letting their semantic feature space evolve during training, leading to faster convergence and better image quality.

MLLMs can get a surprising visual reasoning boost from a simple trick: adding just a dash of visually grounded self-supervision to instruction tuning.

By predicting future scene structure in the feature space of a frozen vision foundation model before rendering pixels, Re2Pix achieves state-of-the-art video prediction with improved temporal consistency and perceptual quality.

Franca leapfrogs proprietary vision models like DINOv2 and CLIP, proving open-source can win on performance and transparency in visual representation learning.

Video pre-training can drive autonomous vehicles, but scaling model size doesn't always guarantee safer closed-loop driving.