CNN Visual Guide
code
V1.2 · ResNet-50 visualizer now live

See how vision models
actually see

Animated explainers teach the idea. Real trained models — from a 99% MNIST classifier to a 76% top-1 ImageNet ResNet-50 — prove it.

Launch ExplorerRead About
emoji_eventsLetter of Commendation · NUS SCALE Deep Learning capstonetrending_flat

Core Concepts

The fundamental building blocks of spatial feature extraction.

View concepts arrow_forward
filter_center_focus

Convolution

Learn how sliding kernels extract features from spatial data through mathematical dot products.

Learn trending_flat
view_in_ar

Receptive Fields

Visualize the area in the input image that a single neuron in a specific layer is sensitive to.

Learn trending_flat
trending_down

Vanishing Gradients

Explore why deep networks struggle to train and how ResNets solved the signal attenuation problem.

Learn trending_flat
visibility

Attention

See how transformers let every position decide which other parts of the input to listen to, and by how much.

Learn trending_flat

Trained Models

Live inference and internal activation visualization on real datasets.

Simple CNN architecture diagram99.2% Acc
MNISTgrid_view

Simple CNN

A lightweight 3-layer architecture, perfect for learning the basics of digit recognition.

60K params
Dilated CNN architecture diagram88.4% Acc
CIFAR-10layers

Dilated CNN

Explores dilated convolutions for increased receptive fields without pooling layers.

2.4M params
ResNet-18 architecture diagram76.6% Acc
CIFAR-100account_tree

ResNet-18

A residual network demonstrating how skip connections enable deeper feature extraction.

11M params
ResNet-50 architecture diagram76.01% top-1
ImageNet-1krebase_edit

ResNet-50

The gold standard for object classification. Inspect activations on a per-channel basis.

25.6M params
DINOv3 ViT-L architecture diagram97.2% orig · 94.9% clean
Weapons · 9-classhub

DINOv3 ViT-L

Meta's self-supervised vision transformer, LoRA-tuned to nine weapon classes — the attention half of an attention-vs-convolution study.

300M (LoRA)
ConvNeXt-V2-L architecture diagram96.5% orig · 96.9% clean
Weapons · 9-classview_in_ar

ConvNeXt-V2-L

A modern convolutional network fine-tuned on nine weapon classes — the convolution counterpart that ties DINOv3 on clean data.

198M params
lock

ViT — Coming Soon

Vision Transformer integration in progress.