Reference cards for the models that define modern vision — the key idea, trade-offs, and when to reach for each.
Robust zero-shot monocular depth from a single image, trained on massive unlabeled data.
The one-stage real-time detector family; modern versions are anchor-free and multi-task.
A real-time DETR variant that beats YOLO on speed/accuracy while staying NMS-free.
A promptable segmentation foundation model that segments any object from points/boxes.
Detects objects from arbitrary text prompts by fusing language into a DETR detector.
Self-supervised ViT features that transfer to many tasks without fine-tuning.
Latent diffusion that generates high-res images from text efficiently in latent space.
A hierarchical ViT with shifted local-window attention — a general-purpose backbone.
Contrastively aligns images and text into a shared embedding space for zero-shot tasks.
Applies a pure Transformer to sequences of image patches — no convolutions.
Detection as direct set prediction with a Transformer — no anchors, no NMS.
Compound scaling of depth, width, and resolution for accuracy-per-FLOP efficiency.
Extends Faster R-CNN with a mask head for per-instance segmentation.
Residual networks that made training very deep CNNs possible via skip connections.
Introduced the Region Proposal Network for end-to-end trainable two-stage detection.
Encoder–decoder with skip connections; the workhorse of medical and dense prediction.