Research Hub

Robust zero-shot monocular depth trained with massive unlabeled data and a strong teacher.

An explicit, real-time alternative to NeRF that redefined neural rendering with millions of 3D Gaussians.

Scaled self-supervised learning to produce general-purpose features that transfer without fine-tuning.

Fused language into a DETR detector for open-vocabulary detection from text prompts.

A promptable segmentation foundation model trained on 1B masks — zero-shot 'segment anything'.

Ran diffusion in a compressed latent space, making high-res open-source text-to-image feasible.

Contrastive image–text pretraining enabled zero-shot vision and powers text-to-image conditioning.

Masking 75% of patches and reconstructing them yields powerful, scalable self-supervised ViTs.

A hierarchical windowed Transformer that became a general-purpose vision backbone.

Showed strong self-supervised representations from contrastive learning with heavy augmentation.

Brought pure Transformers to vision; with scale, attention rivals convolutions.

Made diffusion models competitive with GANs and set off the modern image-generation era.

Detection as set prediction — removed anchors and NMS, seeding a family of Transformer detectors.

Encoded 3D scenes in an MLP for photorealistic novel-view synthesis, igniting the neural-rendering wave.

The Transformer — self-attention replaced recurrence and now underpins language and vision alike.

Added a mask branch and RoIAlign to Faster R-CNN for accurate instance segmentation.

Reframed detection as a single regression, making real-time object detection practical.

Normalizing activations stabilized and accelerated training, enabling much deeper networks.

Skip connections made 100+ layer networks trainable — arguably the most influential architecture idea in deep learning.

The Region Proposal Network made two-stage detection end-to-end trainable and defined the detection playbook.

Encoder–decoder with skip connections — the backbone of medical segmentation and, later, diffusion models.

The adaptive optimizer that trains most deep networks today; AdamW is the modern default.

Framed generation as a two-player game between generator and discriminator — a decade of generative research followed.

Showed that depth with simple 3×3 conv stacks drives accuracy; VGG features are still used everywhere.

The spark of the deep-learning revolution — GPUs, ReLU, and dropout crushed ImageNet and changed the field overnight.

The original convolutional network — established convolutions, pooling, and end-to-end gradient training for images.