AI People Wiki

Builders & open source

Katherine Crowson

Creator of the k-diffusion library and co-first author of VQGAN-CLIP

Katherine Crowson, whose GitHub profile reads "AI/generative artist", is the author of k-diffusion, an MIT-licensed PyTorch implementation of Karras et al.'s EDM diffusion framework with improved sampling algorithms and transformer-based diffusion models. Crowson is co-first author of "VQGAN-CLIP: Open Domain Image Generation and Editing with Natural Language Guidance" (2022), listed with an EleutherAI affiliation; the paper credits Crowson with originating the idea of combining VQGAN with CLIP and with heading the approach's development. Crowson is also first author of "Scalable High-Resolution Pixel-Space Image Synthesis with Hourglass Diffusion Transformers" (2024), and the k-diffusion README documents an hourglass diffusion transformer model type.123456

At a glance

Current role
None confirmed in the sources
Work listed
2
Sources cited
8
Profile reviewed

Open-source work

Crowson's k-diffusion library implements the EDM framework from "Elucidating the Design Space of Diffusion-Based Generative Models" (Karras et al., 2022) in PyTorch, adding sampling algorithms and transformer-based diffusion models. It includes the image_transformer_v2 model type, which draws on the Hourglass Transformer and DiT and uses neighborhood attention at low levels of the hierarchy and global attention at higher ones.23

Research

The 2022 VQGAN-CLIP paper, with Crowson and Stella Biderman as co-first authors, describes steering a VQGAN image generator with CLIP so that images can be generated and edited from text prompts; the paper credits Crowson with originating the idea and leading its development. Crowson is also a co-author of LAION-5B, an open image-text dataset paper in the NeurIPS 2022 Datasets and Benchmarks track.547

In 2024 Crowson was first author of the Hourglass Diffusion Transformer (HDiT) paper, on generating high-resolution images directly in pixel space with a hierarchical transformer. In 2026 Crowson co-authored LoRA-Muon, a paper on spectral steepest-descent optimization for low-rank adapters.68

Work and contributions

  • k-diffusion Open-source code

    Author of the k-diffusion repository, whose MIT License names Crowson as copyright holder; the README describes it as an implementation of "Elucidating the Design Space of Diffusion-Based Generative Models" with additional sampling algorithms and an hourglass diffusion transformer model type.236

  • VQGAN-CLIP Published research

    Co-first author with Stella Biderman; the paper's author contributions credit Crowson with originating the idea of combining VQGAN with CLIP and with heading the approach's development.54

Timeline

  1. Released "VQGAN-CLIP - Open Domain Image Generation and Editing with Natural Language Guidance" as co-first author.5
  2. Co-authored the LAION-5B open image-text dataset paper (NeurIPS 2022 Datasets and Benchmarks).7
  3. Released the Hourglass Diffusion Transformers paper on pixel-space image synthesis as first author.6
  4. Co-authored "LoRA-Muon - Spectral Steepest Descent on the Low-Rank Manifold".8

Documented roles

The sources do not confirm a current role; these are the roles they document.

  • Listed with an EleutherAI affiliation as co-first author of the VQGAN-CLIP paper (2022), EleutherAI4

Companies in their history

Suggest a correction or removal

Profiles cover public professional work only, and each statement cites a source. If something is out of date, missing or wrong, tell us and link a source. People can also ask for their profile to be corrected or removed.

References

8 sources · profile reviewed Oct 21, 2026 · data updated Oct 21, 2026

  1. crowsonkb (Katherine Crowson) on GitHubGitHub · Official page · accessed Oct 21, 2026
  2. crowsonkb/k-diffusion READMEcrowsonkb (GitHub repository) · Repository · accessed Oct 21, 2026
  3. crowsonkb/k-diffusion LICENSE (MIT)crowsonkb (GitHub repository) · License · accessed Oct 21, 2026
  4. VQGAN-CLIP (arXiv HTML version: author affiliations and contributions)arXiv · Paper · published Sep 4, 2022 · accessed Oct 21, 2026
  5. VQGAN-CLIP: Open Domain Image Generation and Editing with Natural Language Guidance (arXiv 2204.08583)arXiv · Paper · published Apr 18, 2022 · accessed Oct 21, 2026
  6. Scalable High-Resolution Pixel-Space Image Synthesis with Hourglass Diffusion Transformers (arXiv 2401.11605)arXiv · Paper · published Jan 21, 2024 · accessed Oct 21, 2026
  7. LAION-5B: An open large-scale dataset for training next generation image-text models (arXiv 2210.08402)arXiv · Paper · published Oct 16, 2022 · accessed Oct 21, 2026
  8. LoRA-Muon: Spectral Steepest Descent on the Low-Rank Manifold (arXiv 2606.12921)arXiv · Paper · published Jun 11, 2026 · accessed Oct 21, 2026