StyleGAN's own dataset record limits its licence to non-commercial use

The process wall

On 12 December 2018, Tero Karras, Samuli Laine and Timo Aila, researchers at NVIDIA, posted a paper proposing a different way to build a GAN's generator. The paper, later marked as its 'CVPR 2019 final version', borrows from style-transfer literature: instead of feeding a latent vector straight into a convolutional pipeline, it injects a learned style at every resolution of the network. The wall it addressed was control: earlier GAN generators produced convincing faces but gave a maker no reliable way to change one attribute without disturbing the rest.

What the documents show

The paper states its architecture enables 'automatically learned, unsupervised separation of high-level attributes' such as pose and identity, and separate control over 'stochastic variation in the generated images (e.g., freckles, hair)', alongside 'intuitive, scale-specific control of the synthesis'. Training a generator to expose those attributes cleanly required a large, consistent dataset, which the authors built for the purpose: the Flickr-Faces-HQ dataset repository states its 70,000 images 'were crawled from Flickr' and published there by their own photographers 'under either Creative Commons BY 2.0, Creative Commons BY-NC 2.0, Public Domain Mark 1.0, Public Domain CC0 1.0, or U.S. Government Works license' — a photographer's copyright licence, not a release from the people photographed.

Authorship and limitation

That distinction is the mechanism worth naming plainly: a Creative Commons licence on a photograph governs what its photographer permits done with the picture, and says nothing about whether the person shown agreed to appear in a face-generation training set. NVIDIA's own dataset repository acknowledges the gap, stating a commitment 'to protecting the privacy of individuals who do not wish their photos to be included', offering an opt-out tag and a removal contact, and stating plainly that the dataset 'is not intended for, and should not be used for, development or improvement of facial recognition technologies'. The code repository itself carries a Creative Commons BY-NC 4.0 licence from NVIDIA: the official weights and code are released for inspection and non-commercial reuse, not to build directly into a commercial product.

What survives

Editorially, what a maker evaluating any face generator descended from this line can check is whether its training images carry a photographer's redistribution licence, a subject's consent, both, or neither — the three are not interchangeable, and only a dataset's own documentation answers which applies. The disentanglement claim is also checkable on its own terms: a generator that truly separates pose from identity should let a maker hold one constant while varying the other, not merely produce plausible faces at random.

• Read the FFHQ repository's licence list against a single downloaded image's stated source licence.
• Compare the paper's stated separation of 'high-level attributes' from 'stochastic variation' against two generated faces that share a pose.
• Ask whether a face-generation tool documents its training images' licences, its subjects' consent, or neither.

The paper's own framing was architectural, not promotional: it names a generator design, not a finished product, and the dataset's opt-out and non-commercial terms are the parts of the record that most directly bound what a maker could do with the result.