This page is an information source for members (and others) who are interested in how developments in AI will affect photography and photographic competitions. It is a set of links to systems, articles, videos and opinion pieces on AI and related topics. This area is changing very quickly so you should be aware that information provided here which was valid at the time of inclusion may be out of date by the time that you read it.
There are thousands of articles on AI and photography and our idea here is help members develop a broad undertanding of the technology and ethical issues around that. If you have suggestions of articles for inclusion here, please mail email@example.com.
General articles on photography and AI
Other lists of AI and photography articles
Joe Houghton’s list
Generative AI systems
Generative AI systems are systems, such as ChatGPT and Midjourney that can generate articles or images based on an very large language or image models that are created through the statistical analysis of immense volumes of data, which has generally been scraped from public-facing web sites. The texts or images that are generated are sometimes practically indistinguishable from human-generated content.
Image generators are generative AI systems that can generate photo-realistic images based on natural language prompts from the image maker such as ‘Scottish Highland scene with mountains in the background and a waterfall in the foreground’. Famously (or notoriously) an image generated by DALL-E won a Sony photographic competition in 2022 (the photographer refused the award).
AI image generators are mostly based on a technique called stable diffusion. The following article explains how stable diffusion works. It’s quite technical but you can skim bits to get the general idea.
There are several large-scale image generators available. Some have a free option that you can experiment with. However, Midjourney and Dall-E2, which are reputedly amongst the most advanced, do not have a free trial option. Adobe Firefly is available to users that have an Adobe photography account. These systems are based on natural language prompting.
Most image generators rely on users inputting natural language prompts describing the image that they want. This can be quite difficult and time-consuming. NVIDIA Canvas takes a different approach and allows the user to create a rough sketch of the image that they would like to generate.
AI-assisted editing systems
AI assisted editing systems are photo editors that incorporate AI technology, including but not-only generative AI, to simplify editing operations. There are lots of general-purpose options such as Topaz AI and, at the time of writing, Photoshop has just released a beta version that incorporates generative capabilities. There are also more specialised systems for portrait photography.
An (incomplete) list of AI assisted photo editors:
A comprehensive review that concludes that there are only marginal improvements in the latest version over the first version of Topaz AI.
A review of the most widely used photo editor, with a good sumary of the AI-supported features in Photoshop. He concludes the generative AI facilities haven’t yet reached a professional standard but it won’t be long until they do.
Note that US copyright law is distinct from UK law. My understanding is that US law does not allow computer-generated works to be copyrighted. This could be an issue for photographers submitting images to US competitions or salons.
Discusses legal issues around copyright of AI-generated images. Based on English law. I guess Scots Law will be similar.
Members’ opinions and experiences
The opinions that are included here are the personal views of the writers. They are included as contributions to the debate on how photographic societies should respond to developments in AI and do not represent the official position of Edinburgh Photographic Society.
Initial experience with DALL-E (David Stevenson)
Experiments where I asked Firefly to generate images similiar to ones that I had taken. It didn’t do very well. My conclusion is that it’s best if you can be specific and the system has no knowledge of context.