Generative AI – My journey creating Sino Amazons.

A quick product description.

By now, you have probably seen a few tasters and hints about an up-and-coming product from 9ja Cosmos – Sinosignias.

In summary – this is a product of two components – A Sino Amazon (Tokenized Artwork) (Think NFT), is a member of a Female Warrior Sisterhood that terrorised the East Asia Seas 1100-1700 AD.

A Sinosignia is a symbol that resembles (but isn’t) an eastern language character. It comes as a Web 3 TLD (Top Level Domain), minted to the Handshake Blockchain.

Each Sino Amazon has their own unique Sinosignia. It represents their commitment, oath, bond, and seal.

As a set, each component has a ‘multiplier’ effect on each other in terms of security and utility.

In short, as a pair, and in alignment with the product narrative, the two parts are impossible to steal, and have more uses than either would on their own. This is what gives them their value.

But I want to leave the product there, because this is about my journey with generative AI since I started trying to do the Sino Amazon artwork myself.

Beginnings.

I’m not much good at art. What happened was I got this idea while securing lots of country centric domains to extend the .9jacom program. I ended up inadvertently picking up a lot of domains cheap that started ‘vlad’ and ended in something that rhymed with ‘the impaler’, and they sounded really jokey – like ‘vladthepondsailor’.

I head the guy that did artwork on the ‘bored apes’ was going out on his own to do something on a theme called ‘Sneaky Vampire Syndicate’ and I wanted to get in touch to see if some partnership could be done to combine the tokenized collection and the crazy ‘vlad’ domains as a joint offering.

All digital art collectible ‘NFT type’ projects were ‘in the toilet’ and there had been some high profile wallet compromises involving the few that were maintaining some value, so in the segment downturn, a product that offered more, and would be much harder to steal, seemed a good idea.

I made a few enquiries and asked around, but its like nobody could find the guy.

By this time though, I became more intrigued by the concept and wanted to realize it some other how, rather than it having to be about SVS and my ‘vlad’ domains specifically.

I’d seen some artwork on Pinterest which had women warriors in ancient east Asian attire though the women looked mixed race between east Asian and sub -Saharan. At the same time, while I was logged in on a node,  a large number of domains of what seemed to be one eastern language character appeared in the domain auctions. It seemed the auction author had forgotten about them (yes, that actually happens!), though when I checked, the characters had no meaning.

The domains had only nominal bids on them, and I put very slightly larger bids on myself, which didn’t get challenged. That is when I got the idea of doing the artwork myself rather than relying on a partner. So I set about writing a narrative that could tie some variation of these images I’d seen on Pinterest and the newly acquired domains.

Then the task was going to be about creating a range of images, and wining sufficient domains that could pair with each other in compliance with the narrative. I reckoned I needed about 1000 sets, so, a lot of hard work. I started in October 2022.

Early efforts

In October, My first efforts to create something in line with the Sino Amazon scope, were on the first version of DALLE E. It didn’t come out terribly well. By November, it was possible to use DALLE E 2. I created some images on the free allocation, but the gap between what I was aiming for and what I got was bigger than the Pacific and the Atlantic added together.

By mid November I began trying Stable Diffusion based providers. I also gave Midjourney a try. I used all the free credits for the MJ trial, but the cost of heavy use was high.

The base engine of SD is open source, so many providers use it, but their implementation seems to be different and it impacts on the quality of results. I tried DreamStudio, Fotor, NightCafe and a few others that online descriptions say ‘seem’ to use Stable Diffusion, though the sites actually don’t say.  I found DreamStudio to be the best overall implementation of SD, but Midjourney still had a superior ‘finish’ quality compared to others. There were some ‘organic’ aspects to DreamStudio however, which had merit, which I will describe in a moment.

I built up a strong repository of DreamStudio content, for which I needed an artist to come in as a partner and do some clean-ups. They weren’t yet up to a marketable standard. I had some great discussions with some folk on LinkedIn who seemed very enthusiastic about getting involved in the project, impressed by the wider aims of 9ja Cosmos. Two people in particular were in detailed DMs with me almost every day.

However as time went on, they never became technically involved and also drifted away from regular LinkedIn presence. Perhaps they had wider issues than a change of heart about the project.

I eventually bit the bullet, dumped some subscriptions to things where I could, and took up the flagship product in the Midjourney options.

The Hollywood Effect

I wasn’t born into a household with a TV. In an age when people can stream video on a mobile phone, this probably seems incomprehensible to many. I can’t remember when a TV came exactly, but it was a black and white TV screen, and the national station had only one channel.

Public funds were very tight, and there was no domestic film industry, only the news and a few domestic current affairs shows. Anything feature length was from abroad, and a minimum of 10 years old before we got it.

When we could see international content, a frequent staple  I could remember were westerns of a type that were called ‘Cowboys and Indians’. Where I was at that time, Native Americans were known as ‘Red Indians’. As a small kid I never understood this. They were a shade of grey. It was a B/W TV. Nobody locally had a colour one. I was this sort of kid that always asked adults loads of questions and would then reason with them why I thought the answer I was given was flawed. I would keep going with the line of questioning until I thought the modified answer I got was reasonable. Sometimes I didn’t get to an end point and got a slap on the back of the head from a random adult. This however never dissuaded me from my questioning techniques.

I was told that in the studio, where they make the film, the Indians were probably red, but we had no way of telling, because the B/W TV can only show shades of grey. We had no Native Americans living locally, so I couldn’t test the thesis, and had to (grudgingly) accept this as something neither proven nor disproven. Armed with the ‘shades of grey’ ‘excuse’, adults now became emboldened to provide me with this answer to many queries I had about things I saw on TV.

Then I had the ‘gotcha’ moment. Most of these ‘Red Indians’ were painted as the bad guy, and were ‘extras’ and usually died in the films. But then I noticed in another film that a certain ‘Red Indian’ was an ‘extra’ as a bad guy in a group of outlaws lead by Gene Hackman, who all got shot up by John Wayne. The same ‘extra’ cowboy, previously an ‘extra’ Indian,  was a white guy.

So I now started asking why did they have to use someone who isn’t a ‘Red Indian’ to be one in a film, since the film was made where they come from?

This was my earliest exposure to – ‘The Hollywood Effect’.

I’m not six years old anymore, and both visual technologies and media through which to educate oneself on far distant lands has exploded. White guys with feathers on their heads and a bit of fake war paint won’t fly anymore but in some other ways ‘The Hollywood Effect’ still exists.

Casting can take someone with fairly soft and malleable features, say Halle Berry, off to Costume and Make-up and in one film, she is an open market stall owner in Luanda, in another, she is a crown princess in Riyadh, in another she is a local civil rights activist in Myanmar. Reality check is that indigenes of these countries look nothing alike, and authenticity should demand the same actor couldn’t possibly fill all these types of roles. To be believable, the best fiction needs to exist in an envelope of fact. This is how escapism becomes marketable.

Strong Points and Weak Points of different AI engines.

Midjourney has by far the best finish quality of any of the AI engines I have tried. But my experience was it is heavily limited by ‘The Hollywood Effect’ , particularly when it comes to imagining authentic and organic Amazons of ethnicities mixed between the countries of their operating region, and other places much further away.

Everything started to head towards a convergence of facial features which started to become common traits irrespective of additional features such as hair style and skin tone. The more variations I created, becoming further removed from the first generation root, the more pronounced these homogenizations became.

Irrespective of the overall drive to have this sense of mystery and pluralism about Sino Amazon origin, there is a sort of organic and authentic realism in what I’m looking for, and I can tell influence from ‘The Hollywood Effect’ when I am looking at it

A Sino Amazon may be an operator in many cultures but is a mistress to none.

While some of the other AI engines give less of a slick finish, they seemed to be able to produce more faithful results from origin influences.

I included some blending, and the use of images from other AIs in prompts for MJ, as well as some manual techniques, which brought improved results. Some semblance of ‘The Hollywood Effect’ still survives in some final product.

Unfortunately, MJ also has one of the most aggressive community standards, and frequently creates false positives for things that wouldn’t bat an eyelid in a PG13 rated film. It seems particularly sensitive to breast cleavage lines that would be obviously apparent in any two piece bathing suit.

I have a lot of raw images good as anti Hollywood Effect’ modifiers for use in blending or prompts, but MJ refuses to accept them on standards compliance grounds. Sometimes MJ even refuses its own output!

‘The Hollywood Effect’ also seems to have an obsession with ‘classic’ perceptions around beauty and body shape. This is also one of the homogeneity drivers I suspect exist in MJ. After all, the more attractive the content people receive, the more likely they are to continue subscribing. Real Amazons aren’t going to look like models in period costume. They are going to come in all shapes and sizes.  Efforts to include references to body shape as a legitimate effort to bypass that aspect of ‘The Hollywood Effect’ often fall foul of MJ’s community standards compliance.

Being Prescriptive

The more outcomes you decide as being clearly either right or wrong, the harder it is to get a good result from AI. AI doesn’t like lots of absolutes (rights or wrongs), it likes a lot of equi-valid versions of ‘different’

The easiest things to do with MJ (and AI in general),  are head shots, head and shoulder shots, strange demonic, alien or abstract stuff. This is because the AI has a high probability of success. The prompt provider has little by way an AC (Acceptance Criteria) or maybe none at all, beyond a vague idea that it ‘must look cool if I post it in online media’.

The more prescriptive the creator needs to be about the output, the harder it gets, as more results will fail.

The subject’s interaction with an object is a common point of failure, because it must be realistic. This is particularly true of weapons. Lots can go wrong with the placing of weapons or the grip on weapons. Trying to inject motion creates more complexity, and therefore more chance of failure.  I try to ‘encourage’ the output to feature some form of a feather weight ultra-sharp blade, because a fluid grip, while in motion, is plausible, a bit like playing table-tennis.

On the other hand, a ‘Bastard Sword’ of the type typically needing a double-handed grip by 400lb medieval crusader in chainmail, just looks laughable when pictured, held with just thumb and forefinger by a sub 150lb woman as if she is carefully getting rid of a shard of glass.

Frequently, results have blades suspended in air, penetrating the subjects body, or simply not present at all. Contortions of the subjects body can also portray a result as implausible.

Period authenticity is another prescription. There is some latitude with the Amazons themselves, as the narrative describes them in a cauldron of local folklore with myths around demonic skills, necromancy, cybernetic implants, alchemy, shape shifting, time travel (non exhaustive list). The environment however is coastal land and seas stretching between modern day Indonesia and Japan, from 1100-1700 AD.

Non period elements of the background, such as evidence of electric power, self-propelled vehicles, air transport, and post-period manufacturing, can happen really easily, and invalidate an otherwise great result.

Appraising other people’s work

I get asked a lot, actually not by Generative Art Creators, but by other people, mostly on LinkedIn, what do I think of…. and they will mention the art of some other creators.

Of course, I can’t really tell, because, particularly with Midjourney, the talent isn’t so much in the image quality but in the success at being prescriptive.

I’ve given some ideas about the sorts of things that are difficult or easy to achieve, but I really can’t tell unless I know the creators Acceptance Criteria, or Success Criteria.

This is on a spectrum from 0 to Infinity almost. It is fairly impossible to get Generative AI to create art around a sequence of numbers, letters, and/or symbols.

The most difficult action still with period weapons, for me would be a bow and arrow still,  resembling shots from maybe ‘Robin Hood’ (2010). ‘The Hunger Games’ (2013) also features archery, but the tools are far more modern.

Archery is more complex than spear, blade or staff type weapons, as the subject has to interact with three separate parts, the grip on the bow, the bowstring, and of course, the arrow. Even getting the AI to faithfully create a believable image of a subject walking passively with a bow in one hand is quite difficult, far less for an action still.

Non period archery tools would be less than ideal, but not a show stopper, as the Sinosignia narrative allows Sino Amazons amazing mysterious attributes that transcend time and space.

Happy to take questions or receive observations that don’t compromise 9ja Cosmos IP.

 

 

Thank you to those who have come out to support the new 9ja Cosmos page

We do however need the strong support of ‘Tekedians’ to move the needle forward, so more followers for the page are urgently needed.

9ja Cosmos is here… 

Get your .9jacom and .9javerse Web 3 domains  for $2 at:

.9jacom Domains

.9javerse Domains

Visit 9ja Cosmos

Follow us on LinkedIn HERE

 

All reference sites accessed 01/08/2023

imdb.com/title/tt1392170/

https://en.wikipedia.org/wiki/Robin_Hood_(2010_film)

nftevening.com/sneaky-vampire-syndicate-partnerships-are-an-incredible-start-to-2022

zapier.com/blog/ai-art-generator/

hunter-ed.com/newsouthwales-bows/studyGuide/Parts-of-a-bowstring/203601_144469/