Another paint over for today:
4 Likes
Gorgeous! How are you combining the two algos?
1 Like
I found some of their posts from Discord interesting too - we’re seeing a lot of the same debates/arguments come up, and people often fall on one side of the fence or the other. I like that he called out the detractors for asking how common it was for an artist to state their methods.
If nothing else its fascinating to see what ‘art’ means to everyone, how they derive value from it etc.
AI art seems to really bother some people - and I’m not sure even they fully understand why.
1 Like
its implemented wirh the --test and–testp variables
2 Likes
At this point I have to go back to my initial thoughts that I have already expressed a little earlier in this thread, since I see an important leap happening here that should not be left uncommented:
As much as my enthusiasm towards the more abstract, dreamylike and highly stylized first versions of midjourney grew over time, even leading to implementing AI generated images partially in my actual works and experiments, I must say that this new step up by implementing Stable Diffusion did a very extreme change.
The last images that have lately posted speak for itself: Photorealistic AI generated images are heavily based on racist and sexist biases.
For example: I tried to get images of Iranian people. They alwalys turned out to be muscular bearded men or (partially) veiled, perfectly beautiful women, who oddly had Indian religious symbols somewhere implemented in the image.
When I typed in “Maori” or “Aztek” I got the images that I have posted here, and after sharing some of these in my network, where I have a lot of friends coming from South- and Middle America, some of them got offended, even reaching out to me what these images are about and who did these “photographies / renders” (I did not mention that these were AI generated).
There are neither indigenous looking businessman, nor women with more “natural” bodies (which equals to bodies that don’t look like topmodels) in the model databanks. And at one point last week midjourney had to take down the access to the photorealistic Stable Diffusion models, since they pretty much oversaturated their servers with NSFW content like a lot of pornography, nudes, gory stuff and even child abuse imagery…
One might argue that working more on prompt crafting will improve that, but this is also exactly my point: People, even whole industries, will very surely adapt and implement these tools, Adobe for example has already announced to implement a textprompt to image filter in the next versions and opened it to beta testers.
And people will not put the effort and time into these tools to generate more unique and original content, since the very reason for using these tools will be saving time and money.
My point is that these tools are very much continuing and worsening these biases, if noone intervenes, which seems to be a quite lost battle, since setting up such models, learning processes etc. require huge datasets and extremely lots of GPU Processing power via cloud computing, which happens to be one of the most expensive services to have nowadays.
I will definitely go on exploring and working with these tools, but I will also look forward to avoid these very cliché and destructive biases in the near future, since I do not look forward to feed these models with more content like that.
That being said I see the very same problems with AI generated music, but with another touch to it: The generations always sound like the most awfully boring muzak in Youtube videos, having close to zero dissonances or originality to them. There was actually only a single example of AI generated music that interests me, which happens to be the “Blossoms” release by Emptyset, since they tried to rather synthesize sounds out of a model that they created in a small team and only fed it with field recordings and sounds based on natural materials, such as wood. “Blossoms” caught my attention since it was so obvious that the machine behind these processes was very much present in the sound, and its struggle to come up with results of something that it purely could not understand (yet), such as fieldrecordings, resulted in these very harsh and outworldly soundscape. I would rather look forward to these very niche apporaches to AI generated media in general in the near future.
6 Likes
Good post and it’s a really important topic, and not one restricted to AI. Different kind of example but this sort of bias exists even where there is human intervention, in pretty lofi situations. One example that springs to mind is how cameras, and more recently, camera software, are trained on white people.
2 Likes
This is just a little glimpse to what AI bias already is causing in other areas
1 Like
You gotta be kidding! This is straight from midjourney? 
yep, but it is based on their two “test” parameters, which is actually a customized Stable Diffusion model under the hood
Still very impressive! The quality of the images, the detail all blows my mind. Thanks for sharing 
1 Like
You make a fair point. If you read through the twitter account of https://twitter.com/EMostaque the guy behind SD you will see that one of the reasons he has chosen to open source SD is to open up the technology to allow anyone to train the engine with their own more ‘representative and inclusive’ material. It of course opens up the question of how do you fairly and accurately represent the visual world in all of its myriad complexity - but that’s for another day.
2 Likes
