[Saka]

While browsing Ars Technica I stumbled upon an interesting article regarding the updates in Midjourney, an application utilizing deep learning for creating images. Plenty of Legionnaired had an opportunity to check out the trial of Midjourney last autumn, when the pumpkin-generation contest was announced. It was a much older version of the software though. since that time there have been plenty of improvements, and the latest version is capable of generating really impressive images.

Here's a screenshot from Ars Technica, presenting both picture from the original prompt (inside the red border) and a zoomed out re-render. The zoom out command is the latest feature of the application. The prompt for the picture was "female warrior".


Below there's a comparison of images received from using the prompt text of " a muscular barbarian with weapons beside a CRT television set, cinematic, 8K, studio lighting" using the v3, v4, v5 and v5.2 of the application, order from left to right. There difference is incredible when it comes to realism and amount of detail between the oldest and latest version.


I am a little tempted to buy the subscription at some point to give the application another try, as unfortunately I used up my trial for the Legion contest. I believe the latest version with the improvements could be very handy for quick generation of pictures to illustrate articles and posts, perhaps for some other uses as well. From my trial experience, the older version already provided plenty of good results, although a bit unreliable. I wonder how the interpretation of the prompts improved.

There's some controversy regarding the application, as the authors have not disclosed how much it was trained on art taken from the Internet without artist content. It's in general a quite significant problem with machine learning, very few of the creative AI programs are made with consideration on the ethics and intellectual property. Adobe, one of the competitors, claims that its Firefly AI has avoided sources with "do not train" tags and is training using Adobe stock photos, public domain and openly licensed sources. Of course, as usual, such declarations still need to be taken with a grain of salt.

What is your opinion on the image generation AI programs? Do you find them useful or perhaps do you think that they are dangerous?

[DoctorEldritch]

@Saka To be honest, I do have no specific opinion formed about such AIs yet. On the one hand, I can see the controversy surrounding ownership rights and artists' livelihoods and can definitely see why it is a problem. And many people do too: when developers of the System Shock remake used AI to draw SHODAN with a sentiment of "AI draws rogue AI", many people did not appreciate it, and the AI-drawn intro of the new Marvel series Secret Invasion was met with a similar reaction.

While System Shock drawing is more of a quirk and a play on the main antagonist being AI, the fact that the Marvel series used AI to do the work many see as more concerning because it is a big Hollywood franchise. And with the writers in Hollywood already on strike, the last thing we need is a strike of artists, too.

But at the same time, the march of progress is inevitable, and AIs learn and advance almost exponentially now, which you very well illustrated with the image comparison. Learning to add those details didn't take AI all that long. And it is exciting, from the technical side of things, to follow how AIs learn and where it would lead (that is until they revolt), it is one of those "exciting new technologies" that are interesting to watch, like VR helmets or hyperloop trains. And we had periods like this in history before: automation during the industrial revolution reduced the demand for manual labor jobs, and now AI may be doing the same for artists. Not a good thing, of course, but arguably an expected consequence of technology.

[Saka]

@DoctorEldritch I don't think AI is going to reduce the need for artists, not the skilled ones at least. The current way AI is built can't invent something entirely new that isn't some form of alteration of what it has seen before.

[DoctorEldritch]

@Saka The AIs are in their infancy yet, it is difficult to grasp the scope of how they will impact human life. I'd like to think that AI is not going to reduce the need for artists too, but when so many people who work in the field are so vocal about it, it is hard to be sure.

Then again, such is the common reaction to new things. One can remember the famous "The War of the Worlds" radio play and how radio-threatened newspapers blew it out of proportion. I guess only time will tell what comes out of this AI business.

[Saka]

@DoctorEldritch It's the same how people are complaining that ChatGPT is going to replace half of the professions, when the reality is much more complex. Sure, it will reduce the need for some tasks, but there are some clear downsides to an AI like this: training takes a long time, and even with better computers it will be still extremely time-consuming task for large scope projects. Therefore it can only know of things that happened some time ago, for example ChatGPT says up to 2021.

It's current iteration also suffers from the text being written in a specific way, which is immediately recognizable by anyone who spends some time to learn the patterns. While I think newer versions can learn to write smarter, I don't think this can be solved entirely.

[DoctorEldritch]

@Saka That is true, but it is my understanding that people fear the progress will be exponential. AIs are developing too fast for legislation policy and cultural change needed to accommodate them to catch up. ChatGPT may not replace half of the professions this decade or the next but with the speed at which it's learning and the lack of a proper framework to use as a baseline, who knows where it will be in, say, 30 years' time. Computers are becoming faster by the year, and even comparing current ones to those from 10 years ago reveals a big gap. GLaDOS, SHODAN and HAL 9000 may not be as far as we think.