Justin Johnson

Justin Johnson is a computer-vision researcher and cofounder of World Labs, alongside Fei-Fei Li, Ben Mildenhall and Christoph Lassner. By November 2025, the company had recently launched Marble, a generative world model that creates editable 3D environments from text, images and other spatial inputs.[2][1]

Before World Labs, Johnson's research addressed how neural networks describe image regions and transform images. He received his PhD from Stanford University, advised by Fei-Fei Li, who also collaborated on his dense-captioning and perceptual-loss papers.[4][5][6]

Johnson, Andrej Karpathy and Li Fei-Fei introduced dense captioning in DenseCap, submitted to arXiv on November 24, 2015. The task combines locating salient regions within an image with describing those regions in natural language. It generalizes object detection when descriptions are single words, and image captioning when one predicted region covers the entire image.[5]

Their Fully Convolutional Localization Network combined a convolutional network, a dense localization layer and a recurrent neural network language model that generated descriptions. The architecture processed an image in a single forward pass, required no external region proposals and could be trained end to end—bringing localization and language generation into one jointly trained system.[5]

In Perceptual Losses for Real-Time Style Transfer and Super-Resolution, submitted to arXiv on March 27, 2016, Johnson, Alexandre Alahi and Li Fei-Fei proposed training feed-forward image-transformation networks with perceptual losses. Rather than relying on per-pixel differences, these losses used high-level features extracted from pretrained networks. The method combined a feature-based training objective with feed-forward inference.[6]

For style transfer, the authors trained a feed-forward network to solve the optimization problem previously proposed by Gatys and colleagues. They reported similar qualitative results while running “three orders of magnitude faster” than the optimization-based method. They also experimented with single-image super-resolution, reporting visually pleasing results when replacing a per-pixel loss with a perceptual loss.[6]

Johnson taught Stanford's CS231n in Spring 2017 alongside Fei-Fei Li and Serena Yeung. The course emphasized implementing, training and debugging neural networks for visual recognition. He was also an instructor for Spring 2026, alongside Li, Ehsan Adeli, Zane Durante and Tiange Xiang, with hands-on assignments and a final project focused on practical deep learning for computer vision.[7][8]