Lucid Feature Visualization

Explore neuron, channel, and class features in InceptionV3 and an adversarially robust ResNet-50

Loading model...

Model

Adversarially trained models can't be moved by imperceptible pixel patterns, so their visualizations show recognizable objects

Layer Selection

Browse the model's internal layers, from early edges and textures to late object parts

ImageNet Class

/ -

Choose a final ImageNet class index to maximize at the model output

Objective

Target the model output directly by maximizing a final ImageNet class activation

Optimization Settings

Advanced Options