Webmodels Lena [portable] May 2026

The technical community fractured:

In 2018, Nature and the IEEE officially discouraged the use of Lena. Computer Vision and Image Understanding banned new submissions using the image. Today’s web models (CLIP, DALL-E, MobileNet) are trained on billions of images from LAION-5B or ImageNet-22k. Lena is irrelevant for training. However, she remains the unit test —the minimal reproducible example. webmodels lena

"Lena is actually a bad test image. It’s over-smoothed, has limited dynamic range, and its popularity leads to overfitting. Natural images (BSDS500, ImageNet) are superior." The technical community fractured: In 2018, Nature and

But the web is growing up. New models are trained on diverse, consented, curated datasets. Lena has been retired to the museum of computing—a beautiful, problematic, and utterly foundational piece of engineering history. Lena is irrelevant for training

This is the story of how a single image defined the engineering constraints of the early internet and continues to haunt the ethics of dataset curation. At the University of Southern California’s Signal and Image Processing Institute (SIPI), assistant professor Alexander Sawchuk needed a high-contrast, high-detail image to scan for a colleague’s conference paper. The lab’s flatbed scanner (one of the first) was crude: 100 lines per inch, 6 bits per pixel.