Computer vision papers explained

Computer vision models turn pixels into representations useful for recognition, segmentation, or reconstructing a scene. Their designs differ in what they assume about images and what supervision they need. These explainers connect architectural choices and learning objectives to the visual tasks they enable.

Start with residual networks, see how the Transformer treats an image as a sequence of patches, then explore learning from missing patches. Continue into segmentation, video representations, and methods for rendering three-dimensional scenes.

Start here

  1. ResNet

    Understand why learning a residual makes deep networks easier to optimize.

  2. Vision Transformer (ViT)

    See how image patches become tokens for a Transformer.

  3. Masked Autoencoders (MAE)

    Learn how reconstructing hidden patches provides a visual training signal.

All Vision explainers

13 papers · newest explainers first

VisionLive
Masked Autoencoders (MAE)
Hide 75% of an image, reconstruct the missing patches, and a plain ViT encoder learns to see.
Read the explainer
VisionLive
3D Gaussian Splatting
Store a scene as millions of tiny 3D Gaussians and a GPU renders new views of it in real time, no neural network in the loop.
Read the explainer
VisionLive
VGG
Hold everything fixed but depth: a uniform stack of tiny 3×3 filters, 16–19 layers deep, that reset how convnets are built.
Read the explainer
VisionLive
TRELLIS.2 (O-Voxel)
A field-free voxel that reads where a surface crosses, so open, hollow, and tangled 3D shapes all fit.
Read the explainer
VisionLive
NeRF
Fit a small network to a scene’s photos, and render it from any new camera angle.
Read the explainer
VisionLive
V-JEPA
Learn video by predicting the features of masked spatio-temporal tubes, no pixels and no labels.
Read the explainer
VisionLive
I-JEPA
Predict the embeddings of masked image blocks, not their pixels.
Read the explainer
VisionLive
V-JEPA 2
Predict masked video patches in a learned representation space, then plan robot actions in that same space.
Read the explainer
VisionLive
D4RT
Probe any point in space and time to rebuild a video in moving 3D.
Read the explainer
VisionLive
Segment Anything (SAM)
One promptable model that cuts out any object, zero-shot.
Read the explainer
VisionLive
U-Net
The U-shaped net that segments cells from a few dozen images.
Read the explainer
VisionLive
Vision Transformer (ViT)
Cut an image into patches and a plain Transformer does the rest.
Read the explainer
VisionLive
ResNet
Add the input back, and 152-layer networks finally train.
Read the explainer