Justin Johnson

Johnson completed his doctorate at Stanford, advised by Fei-Fei Li. A University of Michigan announcement describes his research across vision and language, visual reasoning, and image generation, and records that he co-taught Stanford’s CS231n deep-learning course. It announced his appointment as an assistant professor beginning in September 2019. A subsequent Michigan seminar biography also identified him as a visiting scientist at Facebook AI Research.[4][5]

World Labs names Johnson among its founders. Its September 2026 acquisition announcement says he and Ben Mildenhall will continue leading the team with Fei-Fei Li as it joins AMD; the transaction remains subject to closing conditions. For a practical introduction to Johnson’s educational work, his PyTorch examples repository walks through small neural networks, automatic differentiation, model modules, and optimizers. Readers comfortable with Python and basic neural-network ideas can follow how the implementation changes as higher-level tools take over repeated bookkeeping.[1][2][9]

Learning an image transformation once

Johnson, Alexandre Alahi, and Li Fei-Fei proposed training image-transformation networks with perceptual losses. Instead of judging output solely through individual pixel differences, their method uses higher-level features extracted by a pretrained network. A feed-forward network learns to produce the desired transformation, so it need not repeat a slow optimization procedure for each new image. Their 2016 paper evaluates style transfer and single-image super-resolution. It reports substantially faster style transfer with similar qualitative results to the optimization-based comparison, within that experiment’s setting.[6]

Testing whether a model reasons about an image

CLEVR addresses a problem with visual question-answering benchmarks: a model can exploit regularities in the questions and answers without doing the intended reasoning. Johnson is first author of the dataset paper. Its diagnostic questions and detailed annotations distinguish different reasoning requirements, helping researchers locate specific weaknesses. The contribution is a test instrument as well as a dataset. Good performance on ordinary images can conceal shortcuts; controlled tasks make those shortcuts easier to investigate.[7]

Making three-dimensional research easier to implement

Johnson co-authored the paper introducing PyTorch3D, an open-source library of operations for learning with three-dimensional data. The authors identify practical obstacles such as processing differently shaped data efficiently and making graphics operations differentiable, meaning that learning algorithms can adjust their inputs using gradients. The library includes a renderer for meshes and point clouds. This lets researchers connect a proposed three-dimensional representation to the two-dimensional image it would produce, then use that comparison in learning.[8]