Briefing Paper on A.I. Image Generation
Exec Summary:
Machine Learning / A.I. software now exists that allows businesses and organisations to create photorealistic images, drawings and illustrations in-house for free or at a very low cost.
Business Advantage:
Costs of using stock photos, sourcing basic illustrations and designs can be reduced to zero. Smaller businesses will be more competitive.
Business Threat:
Stock photo production companies may see sales impacted in the medium to long term but they will also see a massive decrease in production costs allowing them to infinitely expand their collections
Business Opportunities:
A new creative economy can be created where people with just an idea can will the imagery to life. Create that poster, that comic, the new brochure by typing a few words. The same way the likes of Photoshop and an internet connection unlocked so much creativity, now people can make art like professionals without having creativity suites.
Background:
In the past few months we’ve seen A.I./ Machine Learning advances so that you can type a few words into a website or desktop app and images will get created for you. Or someone takes an existing image and has it modified or “outcrops” it so the A.I. creates a new world around the image.
How this A.I. software works:
This software uses images and text descriptions downloaded from the Web. Billions of them. This is then put through machine learning/A.I. software that links words to the content of images so that it eventually understands plain english commands e.g. draw me a tree or create a photo of a tree.
Malcolm Gladwell talked about putting 10,000 proper hours into something to get good. A.I. is putting those hours in so you don’t have to. The author of the paper Gladwell cited says he oversimplified and there was also another crucial aspect – having a very good teacher. Who are the good teachers in A.I?
4 main A.I. image systems
Currently there are 4 main image systems but there will be dozens soon.
- Dall-E from Open A.I. which was the first, is web based and invite only. First few images are free. Then you pay for credits.
- Crayon – a clone of Dall-E – Free website.
- Midjourney is in Beta, can be accessed via a Discord web server. First few images are free. Then you subscribe.
- Stable Diffusion – Open Source, desktop version. Free.
Dangers and Deep Fakes:
Dall-E has checks to prevent the creation of deep fakes and creating violent imagery but many of the others do not. It can be very easy to fake a person, a celebrity or a politician and have them endorse a product or movement or show them in a compromising scenario.
Ethical Issues
As this “tech” is based on billions of images that have been scraped from the web, these images are biased toward the white male gaze of course and with pornography also downloaded, what this software considers to be a woman can be corrupted. The likes of Dall E does error corrections by adding in women so searches to decrease all male images as well as trying to remove violent and racist imagery. Copyrighted images are also in this giant set of images. Read and follow the work of Abeba Birhane.
What comes next:
Video creation and editing e.g. a text command can change the colour of the sky or reorder a city. This is already happening.
In the medium term people will be able to conjure up their own worlds by just telling their VR device. Think will be a jump from Fortnite to a world like in Avatar or any elaborate computer game.