Frequently Asked Questions–100k Faces
Update: the Generated Photos platform has now launched, for the latest information please visit our official FAQ.
Thanks for the interest in Generated Photos! We know you have many questions about who we are and what we are building; let us clear those up.
Looking for example photos from the 100k Faces project? Visit our press page.
What are we doing?
We are democratizing creative photography and video. We aim to make creative works both more accessible and higher quality through generative processes.
Who are we?
We are a team of about 20 AI and photography professionals.
- AI engineers
- Photographers
- Make-up artists and stylists
- Retouch artists
- Data labeling team
- Freelance models, image masking professionals, and technicians
- A videographer (video is where we go next).
The founder is Ivan Braun.
Where are we based?
Generated Media, Inc. is a fully distributed company (incorporated in the USA). We have been remote-first from day one and now have employees representing six countries and counting!
Where can I reach you?
Contact us anytime at work.with@generated.photos.
What work was done for the 100K Faces project?
We have built an original machine learning dataset, and used StyleGAN (an amazing resource by NVIDIA) to construct a realistic set of 100,000 faces.
Dataset Creation
Our dataset has been built by taking 29,000+ photos of 69 different models over the last 2 years in our studio. We took these photos in a controlled environment (similar lighting and post-processing) to make sure that each face had consistent high output quality. After shooting, we underwent labor-intensive tasks such as tagging and categorizing.
Software Development
- Machine learning model creation
- Scripting for mass image generation
*We plan to write more about how we accomplished this task in greater detail!
Why do some of these faces look strange?
Generating media can definitely come out weird at times! Some faces will have strange ‘stuff’ in the background, while others just seem slightly off for no apparent reason. For example, we have noted a common issue with something that looks like a crater appearing in random places. Those familiar with generative adversarial networks have likely seen worse. Our whole process is designed to gradually improve the final results and we are far from done yet! We find these stages of growth quite fascinating. Take a look at this early sample to see where we started:
We’re currently busy behind the scenes selecting out the highest-quality generated images. It’s a relatively simple classification task, but it does take time; you’ll see more from us soon!
‘I’m concerned {…} people aren’t represented well enough’
After running a studio for the last 2 years, we have learned it can be difficult to find interesting, diverse models that are also willing to shoot stock photography. We are fascinated by the range of personalities we have captured so far and will continue to push that envelope going forward.
Can I use these faces for free?
Yes! They are can all be used for free as long as a link back to generated.photos is placed on the page, post, or project. Currently, faces made available for the 100,000 faces project are available for non-commercial use only.
Update (Dec ‘19): Licensed commercial use of AI-generated faces is now available, please visit Generated Photos for more information.
How can I spot a generated photo?
The most notable errors occur in places that lack adequate training data. Pay special attention to images of uncommon events or poses. Items that can vary greatly such as clothing, hairstyles, or accessories are also especially difficult to realistically generate. For photos of faces particularly note:
- Around the ears: hair can cover ears in many unpredictable ways
- Headwear: we lack multiple images for each specific type
- Collar area: generated clothing varies greatly
- Unusual expressions, such as a broad grin. Back teeth are not often exposed and can appear distorted.
- Backgrounds: with so many potential environments they tend to vary highly in quality and realistic nature
How about marking the generated photos?
Specifying images as generated in situations where there are legal or ethical concerns is perfectly reasonable.
The reality of the current situation is that we may safely assume virtually any image has been photoshopped, beauty-filtered, generated, or otherwise manipulated. Even an out-of-the-box iPhone makes skin look like a barrel of silicone foundation!
Regardless of the process used to modify an image, we (society) need a technology that certifies an image has remained unchanged. Crime scene photography, product shots, passport photos — we need a technology that confirms the final shot is the result of the light passing through a lens, without any manipulation.
What are some impacts generative media will have?
In short: media will change forever.
Lookism goes away
Conventional beauty is changing. Our systems don’t understand or care about a created person being ‘beautiful’ or ‘ugly’. They care about whether a person looks human or not. With generative media, it is possible to build any combination of physical attributes. We will no longer have to rely on a few supermodels to represent the standards for the entire world.
Media will become more personal
Generative media is not all ‘fakes’. The same technology used to generate random unreal people can also be used to recreate people from the real world.
Imagine Netflix being able to synthesize your family members so that they can have small parts in their next hit series. People and situations from your life will be used to deepen your connection to media. We believe that technology is not far off. While it does open the door to a myriad of potential situations, we believe that the benefits from generation vastly outweigh the potential issues.
Cult of celebrity fades away
Why should we dream of meeting someone we don’t know? Why don’t we dream of meeting our family the same way? What if watching a movie suddenly seemed incredibly personal? These are the types of questions we are working to answer.
Reduced barriers for creatives
A lack of resources often prevents newcomers from entering the media industry. You can’t produce a Netflix show without Netflix money. With generative media, a single creative person can create a whole movie just sitting at their computer. This opens up markets across the world to producing content that would previously have been impossible.
Verification of authenticity
We will not produce documentary photography or video.
News footage must be documental, as does a family archive or a photo from a scene of a crime.
Generative media must be kept independent from documentary media. We encourage technology to be built that certifies authenticity — a rubber stamp analog for the 21st century.
