Dominating the Stock Photography Industry with AI Generated Faces for Stock Images

The generated.photos project uses Generative Adversarial Networks to create a large dataset of faces for use as stock images; what will this mean for the future of copyright claims?

[Summary]

Generated.photos is working to not only provide a diverse set of human-centric stock images for use in marketing, but also may be helping to solve copyright claim issues arising from unlicensed uses of likenesses in stock images in promotional materials.  Beyond the free offering of 100K images of generated faces, they also have an API to create faces given a number of parameters to be fed into the GAN like gender, ethnicity, age and even expression/mood. While not stated upfront, a method like StyleGAN would be the engine behind this.  Some photos look quite convincing, but others have clear residual artifacts that indicate they’re generated, so it would be unwise to use this in any automated process. Generated.photos is not the first company to make a tool like this: https://thispersondoesnotexist.com/ released something similar, but lacks the UX. 

[Commentary]

It’s surprising that a commercial implementation of this did not come sooner, but it still leaves many open questions.  As discussed in the previous newsletter, facial dataset provenance is a hot topic, but the creators claim that they operated a studio to gather 30K photos from 70 different models that all apparently gave consent for the AI project. However, if there were a single unlicensed face used to train the GAN, would that person technically be able to make copyright claims on every single generated image?  There are tools in the works to determine the impact of a single training data point on a model that may help assign attribution. A GAN learning from images is parallel to a musician that listens to music for inspiration in making new works, however the former case lacks “real” human creativity. These are important questions to ask that bridge the divide between machine learning and the law that have not yet been addressed by the courts.  More questions that arise are how to reduce data-based biases and ensuring that one has a diverse enough pool of seed faces to represent the entire spectrum of human ethnicities. At least a tool like this, with more seed-faces, will allow marketing materials to represent a wider array of potential customers.