The First step of 100,000.
I started initially messing with AI on a little site call NightCafe about the time that Midjourney first had its major release and was the big hot shit for artificial image generation. Prompt engineering was just barely a thing and we would throw random descriptions at the wall and see what cool shit came out the other side. It was a black box, compared the to sausage maker it has become.
The formula's are fairly well understood, pony training, the porn war, censorship and content warnings, legal and moral questions, gatekeeping, profiteering, and the new real race being the step towards video as developer tools in other applications for visual medium are hammered out. The curiosity of the black box that is AI to most users has not been satisfied just yet, but the artistry and churn of images is for the most part coming to a close.
I think there is a return to this first initial call to creativity with FLUX and FLAN since both can and do make effective meme machines: If you can run them locally. That's the critical missing piece of a lot of the new models, even in distillation, the average user must rely on webhosted generation which in turn must have certain safeguards, filters, and of course Cost compared to Kw hours and hardware depreciation.
This once more returns many of us to the beginning that we started with; pay wall. Scraping likes, generating user interactivity metrics for ad revenue traffic metrics, scumified monetization creep, marketing and sales retention practices to cash in on audience capture, redundancy cosmetics, FOMO, and the chase to the cutting edge anywhere possible to stay one step ahead of competition that would threaten the bottom line. Seen that song and dance more than a few times in too many places. It leads to replacement, dissatisfaction, and competition.
Civitai isn't the first nor the last place I've taken first steps in learning the process of nor do I expect it to be eternal, I do not expect any image I generate or post online to be especially impactful or last longer in the minds of the few who see them in the torrential feed of new images. Some are good, some maybe even great, but they are just a few out of the grist of now over 100,000 images I have created, and someday likely to be 1,000,000's of images I generate.
At 100 steps
The balance of getting started towards being a proficient user has now also dramatically shifted. Lora, style guidance, and meta data is the first plateau of baby steps to getting what i would say is user proficiency. Having enough knowledge and in some cases skills, to generate an image to your ultimately desired result. Even if it is just 1girl posing basics. The floor is very low with most model tools to getting to a user proficiency state. I do not think we know yet, truly, where the ceiling is just yet, but the floor and most of the walls to climb are more or less established.
It is at this step we are at our most fire and go and it is our accidents more than our guided direction that yield typically the most interesting results. Like what happens if you put some things in negative prompting that should be contradictory yet produce desired results on occasion, or how to deal with bad features without just an embed or removing the feature entirely in your prompting.
This is also the peak of learning to smash two things together at varying weights and trying to not noise brick your generation. Sometimes it's certain models, lora with negative interaction, or just for what ever reason, the model doesn't have the right kind or a VAE at all, and you need to go download one for the right model type that it is this time, and pray you aren't wasting disk space or processing time on a weight 1 stack of 8 different lora/embeds/etc. This is also usually where you start learning the features of the WebUI of your choice and the quirky bullshit that comes with it, even if now most of them use the Comfy backend.
The other major factor here for the hosted generation users is you have discovered the metagame magic of how to time your posts for maximum visibility and capturing as much free credit towards generation that you can possibly scrape for yourself. With your standard generation taking 4-16 credits, your capped self directed scrape of maybe 250 nets you the chance at making only 50 images a day on average. If your images are timed properly and good enough, you might net a dozen or so extra total images for the next day until you cap out at around 500 credits or so per day if you are lucky or get reaction scraped by the dude with multiple accounts on a bot script.
At 1,000 steps.
There really isn't much that changes here, other than potentially getting bored of doing 1girl or whatever takes your fancy after making hundreds of generations of it. You may finally have made a few that you consider truly and wonderfully exceptional to keep around for a long time. You find new things to either trend chase or have been cruising the new or popular generations for things to spin off of, start looking at styles more than particular characters, either because you have a particular character/etc you want to make images of or just want to see new ways to make what you like look different.
This is also the step that would normally separate a prompter from a technical user, and both experience the first hazards of noise bricking across model clashing in a much more unfounded way and start to realize 'oh, it's this that makes it bad' on a more fundamental level if they are locally generating by this point. A web user typically either starts becoming very conservative with their tools of choice, narrowing to only the most trafficked or popular models/lora, creating a very different trajectory for their behavior, either burning out entirely on generating themselves or starting to pay in one way or another if not moving to local generation themselves.
10,000 steps.
This is a plateau and a steep one as you climb further, much likely changes unless you are stuck on a singular model/lora set and never update your software etc. Typically this is also where I see many people drop off from generating new images, likely having exhausted much of their interests in the gimmick side of personal use.
The bigger issue comes in the industrial use cases at this stage, where such plateaus lead to stagnation and death. Having to learn new systems and balance the breaking points and bugs. This is also where branching paths really open up; Video, audio, 3D model, and steps outside of 1girl prompting for most.
100,000 steps
Now that I am here, I can say the risk of burnout is staggeringly high, seeking new things to make and really challenge creativity gets harder either from the ceiling from the models, the difficulty in maintaining image quality in niche subject prompts, and the need to learn much more technical tools to overcome these challenges.
The costs also start to become much more recognizable as generation time, power use, and mistakes pile up. For a rough comparison, on Civitai a single batch of 4 images at 35 steps with no lora takes approximately 25 buzz, or around 2 cents. so 100,000 images, not counting bricks and failures, is around $500 worth of content before lora, upscaling, or any other tools to improve and fix content.
You also realize that either the hosting of your content is worth it on the site for engagement and generally free storage and even semi-tacitly are rewarded beyond the dopamine loop if you don't buy the shop cosmetics or otherwise don't recycle buzz into the system in some fashion (though to be fair some shit is cool and shinney exclusive shit tends to work and the trouble of extracting 10$ worth of buzz from the site isn't worth the effort.
So where does this lead next?
The biggest thing I know that I do, is have an impact on the perception, visibility, and content feed in general. I dropped a post of 20 images every 5 minutes and every minute during prime reset hour on Halloween. From this storm of content saturation of goth girls, I pushed over the 100k threshold, broke 1200 followers, capped out buzz rewards and received tips of a fairly significant margin.
The biggest part of this event, was my content dominated the feed in general for an entire 24 hour cycle. Goth girls were enjoyed or hated, and the effect of which was I believe a more or less final saturation of the active userbase at the time of posting. That being the case, I likely won't take the time and effort to ramp up to that sheer quantity of posting as the upload and time release system kind sucks.
I am looking for more stuff to do, maybe get back into some video experiments, try some Unity porting tools, maybe some training attempts if I can get Kohya to play nice, and keep a semi steady stream of much of what you've seen already as I add to my pile of over 600 lora in my prompt pool.
If there's something you'd like me to make more of specifically, you have a style lora etc you want me to send into the grinder that is my prompt algorithm let me know.