One of the biggest developments currently going on, technology-wise, is in the field of Artificial Intelligence. The impact AI can have on society can be tremendous as computers take over tasks and the fabrics of society have to be reinvented. Recently a group of tech leaders, including Elon Musk and Steve Wozniak, even published an open letter asking for a pause on the development of AI to have some time to assess the effects and risks of AI on society. It is not my goal to discuss either the technical aspects of AI or the possible moral and societal questions it raises; I will just talk about how AI can be used for image creation of rainwear. So don’t be afraid of having to go through technical terms or philosophical debates here, I will just share my own experiences of testing current versions of AI and how well they do creating rainwear related content.
The bigger picture
It is not the first time in history that a new invention is about to have a huge impact on society. But what is unique about the progress of AI is the speed with which it is happening. Where the introduction of the steam engine, automobiles, planes, and the internet took decades, it seems like AI is going much quicker.
The best way to explain image creation with AI is maybe by looking at the bigger picture and seeing it as the natural continuation of photography. In the earliest days of photography it took a dedicated photographer who invested lots of time and money into equipment and honing his skills to make a living as a photographer. This changed with the introduction of the compact camera and the 1-hour photo-development stores. Now everybody was able to take pictures and get them printed. Still to create truly great photographs you needed top of the line equipment, skills, and you had to travel around the globe for the right models or sceneries. Back in those days there already was “photoshop”, but instead of a computer program it consisted of a pair of scissors, glue, and a lot of patience. Next digital photography came around: instead of needing lots of skill to get good pictures you could now see a preview on the screen or just shoot unlimited pictures of a scenery and pick out the best one once you were back home. And Adobe Photoshop made it possible to digitally improve your pictures after the fact. Still, you needed to get out and actually take the pictures. You needed the model, the setting with the right light, a wardrobe, makeup artists, etc. And this is where AI comes in as the next step, or maybe the final step. With a simple prompt, which is a description of what you want to get the AI to produce, it can make a complete image in the way you want. You don’t need to go to Paris for a shooting in the streets of Paris, you don’t even need a model anymore. You can now simply tell the AI to create a photoshoot “on the streets of Paris with a blonde model aged 30 wearing a blue business suit while the sun is going down”.
Progress
What is most interesting here is the speed with which AI improves over time. Below are my own tests of using the same prompt in different versions of Midjourney. The way prompts have to be constructed has actually changed over time, but this simplified test will do for now to show the progress.
Below four images created with Midjourney version 1, which was released in February 2022. In the prompt I asked for a photorealistic picture of a blonde woman wearing a black Mackintosh raincoat in the streets of Amsterdam.
As you can see the results are horrible. While it does show a “woman” standing in the streets, wearing something black that is sometimes more yellow, it is hardly an interesting picture. In the second picture you can even see the “shutterstock” watermark still in there, as the AI has been trained (partially) on Shutterstock images. Now let’s use the exact same prompt on Midjourney version 2, which was released April 12 2022. So this version was released just 2 months later.
There is some improvement visible here, but the results might only be useful as stylistic pictures used in a CAPTCHA question to make sure you are human. At best you can imagine a child making drawings like this. There are faces visible, but thats the best I can say. On to version 3 which was released in July 2022, just a few months after the previous version.
This is getting better, although in the top left picture it forgot to make it human. The other three pictures are some stylistic artist impressions at best. It is hard to imagine anyone really interested in printing any of these images, let alone if you are interested in the raincoat specifically. On to version 4, which was released in November 2022. The sequential versions are coming out a bit slower now, but as you can see below the improvements over version 3 are remarkable.
Now we are getting 4 images that are actually impressive: I wouldn’t be able to draw that well even though I think I could outperform versions 1 to 3 with a piece of paper and a pencil. My prompt asked for a photorealistic picture, but none of these are convincingly showing real human beings. It feels maybe as if someone took a picture and used a filter or two before posting it online. It is more like a painting than a picture. On to Midjourney version 5, which was released in March 2023. Again the same prompt was used and v5 gave the following results.
Now we are getting somewhere. Besides that the facial features are getting really close to realistic, also the raincoat itself is becoming interesting to look at. It is now clear that there is a blonde woman standing in the streets of Amsterdam, or a similar looking city, smiling at the camera while wearing a black raincoat. The blurry background, which is often used in portrait photography, is probably a bit over the top, but the reflections of the wet raincoat and some small drops of rain still falling makes these images quite convincing and interesting.
Issues and mistakes
The Midjourney version 5 images might be convincing up to a certain level, the AI still has problems with specific details that give away that its intelligence is artificial. From the previous 4 picture I have developed the bottom left one further, giving the following image.
It will be hard to deny that the facial features are absolutely stunning in this picture, although it does look like some filters have been used. But let’s focus on the raincoat here. You can see clearly some drops of rain laying on, what looks like, the black leather material and the reflection of the light looks very natural. There are some tiny issues though, for example the buttons on the coat seem placed at odd spots and one button hole is missing the actual hole in it. And this is the first problem with the current versions of AI:
Logical fallacies
When you start creating multiple images you will notice some strange things are going on in many pictures: from straps or belts that disappear and reappear, to buttons in strange places, or things that make no logical sense. A good example of this would be the image I created below, showing a lady wearing a protective suit in the mud. She looks very happy and relaxed there, with her hands tucked into the pockets of her suit while smiling at the camera. But why would a suit like this have pockets? It makes no logical sense as they would flood and you will be carrying around lots of mud and water in your gear the whole day. The AI just makes use of thousands of pictures as input where probably lots of them of people standing in coveralls or other regular gear. Now since I asked for rubber gear the AI turned the material to something shinier, but doesn’t understand that protective gear cannot have pockets like that.
Added attributes
Another thing you will notice pretty soon is how the AI is adding all kinds of elements to images. This can be little straps, buttons, or some extra fingers. The main reason for this is that the AI had a limited number of images showing certain things, for example hands, so it is not completely certain yet if a human being has 5 or more fingers. Below an image of a beautiful young lady standing in the mud with some protective gear on, but for some reason she is wearing sort of a belt over her right knee. If that belt wasn’t there, as well as the other item over her other knee, the total picture would be pretty convincing in general although the face seems to be of a slightly lower resolution.
Next another girl wearing rubber chest waders. The picture in general looks pretty convincing, although it could use some improvements with Photoshop around her right elbow. That is until you focus on the right hand, which seems to show 6 fingers even without counting the thumb.
Material details are lacking
Since the focus here is on rainwear, I must mention the difficulty the AI has in producing realistic pictures of rainwear materials. This issue is mostly related to rainwear materials specifically, as the picture below shows that for example leather is being generated much better.
With rainwear it is much more hit or miss, with it sometimes coming very close to the intricate shine of yellow PVC while other times it feels much more like some generic shiny plastic material. Below an image of a young lady in a PVC caving suit where the yellow gear looks decently convincing.
While in the group picture below the suits look much more like general coveralls which have been turned yellow and slightly reflective by the AI. Only with the girl on the utmost right I think the material is convincing around her hips where more reflection is going on.
Brand specific details are lacking
Finally it becomes clear that the AI is producing generic images. When you ask for a person wearing a rainsuit, it will use a generic depiction of a rainsuit that will have reflective stripes one time, multiple colors the next, and the fit will also be all over the place. This way of producing generic images is fine if you are simply interesting in seeing certain materials or types of gear and care less about the details. For example below a picture of three ladies wearing unidentifiable raincoats which perfectly reflect light and show the drops of rain lying on them. But if you would like to see them in a Klepper raincoat, Agu rainsuit, or thick Guy Cotton PVC, then you are out of luck as only generic gear can be produced.
This process of creating images actually made me realize the importance of the brand-specific elements in raingear. Where I absolutely love seeing pictures of vintage Agu rainwear, it becomes a lot less attractive when I notice the gear being worn is a knock-off from another brand. For me these tiny details represent the “heart and soul” of the rainwear, which is also what makes me interested in specific brands and their history. When I see a pair of Le Chameau rainboots I see more than just rubber boots; I also see the early days of rubber production, made-to-measure boots for farmers in France, a production facility set up in Marrakech, and a brand fighting against cheaply produced rubber boots from China.
What will come next?
When you look at the different versions of Midjourney and the quality of images it produced, and notice how quickly these versions came out after each other, it is not hard to imagine what will happen in the near future. The previously mentioned issues and mistakes will most likely be solved soon giving users the opportunity to produce even more realistic images of models in rainwear. The AI will then be able to show different materials more accurately, and I can even imagine brands having an incentive of getting the AI trained on their brand-specific products. Imagine anyone creating an image that contains a pair of rubber boots; it will make a huge difference for manufacturers if those boots have a Hunter or Aigle logo on them from a marketing perspective.
I can also imagine that everyone will be able to either build a model of themselves by entering their personal measurements or getting their body 3D-scanned. With this model you could go shopping online and let the AI pick out outfits for you and show you walking over a catwalk in that specific gear. You will see how different items fit you, what happens if you go a size smaller or larger, and how you can best combine new clothing items with what you already have. While this will probably be led by large online platforms selling millions of clothing items, it is only a matter of time before you can see yourself, or someone else, walking around in your favourite fetish gear produced by AI.
What will happen next, and when, I don’t know. For now AI is mostly a “toy” to play around with by creating different images, but that might soon change. Some artists are already using AI generated images as a base for their artwork and advertisement agencies are working on using AI instead of models, photographers, makeup artists, and photoshop editors. AI certainly gives us interesting times to live in…