- This topic has 58 replies, 8 voices, and was last updated 2 years ago by
.
-
Posts
-
I think people already realise a LLM, or a chkpt, which is a snapshot of the model's parameters, can't independently produce a perfect or flexible result. So even a broadly-trained checkpoint like CyberRealistic with a wealth of knowledge of Prince Charles for example, will only produce a generic image of him. It's why people make Loras or finetune chkpts to modify and add to the "knowledge" of a model. To say that a LLM-based model is incapable or is resistant to learning "new tasks" ignores the fact that people are doing just that every day.
Yes. Refining a template will provide better and more varied results.
I tried AI a few times. The abundance of images I've seen are close enough to be what I'd consider a random generic copy.
For instance typing words like
Pretty Asian girl
Cyberpunk city
Red dressWill probably produce similar results to a dozen other images because of the limited text prompt.
What does the AI think a pretty Asian girl is. I ran around 5-6 images that the girl was very similar in appearance. Do I accept the AIs standard of pretty?
I ran maybe 3 cyberpunk backgrounds that look nearly identical.
It also had a set idea of what a red dress was.
I'm using a less quality online program than flux but I'm guessing the AI is averaging accepted results/terms and settling on a very limited selection of models.
So eventually the computer will say THIS is a Pretty Girl. THIS is what an Asian looks like. THIS is a red dress.
Of course adding in artist styles and themes will change things but pretty Asian girl oil painting abstract might still use the same base image of the girl the AI has settled on.
I'm sure there's a lot more to it than that but the video made me think that why wouldn't the AI eventually say after 20 million base generic add a cat prompts that "THIS" is a cat?
If you've only used generative AI "a few times" I really don't think your in a position to judge its effectiveness or capabilities. Of (fucking) because if you use a very weak prompt like [Pretty Asian girl] with nothing else you'll get a generic character, just the same as the vast majority of Westerners and Daz would vision one, or if you Google the term. Why not complain why Googling "man" produces 90% white Caucasians when the vast majority of men are not Caucasian.
If one wants to be lazy with AI it will reflect that laziness in what it generatively produces. It will throw your own lazy non-specificity back at at you. What the hell expect it to do, read your mind?That's my point. The algorithm for man is a white man. Unless you specifically say old Asian man with dark skin and a wrinkled face you'll get what the AI thinks a man is.
If I were to say Big Cat I'd more than likely get a fat furry house cat not a lion.
Out of probably billions of images the AI has decided by its programing what "cat" is. Not a Persian or a Siamese but just "Cat".
AI won't ask you what breed, color,size,species ect... it just sees what it considers a cat or a man is.
Unless I give AI the exact person/ model I want my Asian girl to look like it will chose the face, age, height, hair color +style, skin tone, ect.. of what it perceives an Asian girl to be. That one specific limited prompt "Asian girl" is what the AI thinks an Asian girl is with little variation.
While I definitely understand complex prompts are (fucking) needed, its not where I'm going with my observation. Every person has a generic idea of what they picture in their mind when they hear the word cat. AI now has a generic picture in it's "mind" of what cat, Asian, girl ect.. are and requires directions to deviate from that. Over time as prompts are reused like "long straight hair" you might get the same long straight hair.
While I'm no AI prompt expert a Google image search now brings up a massive amount of AI generated content. It's near impossible to search Asian girl and not get pages of AI crap that looks the same.
If I suggest to AI I want my Asian girl to look like Lucy Liu does that make it more creative?
If you use an AI like stable diffusion there's many different settings you can use (such as the CFG scale), to do things like allow the AI to be more creative, how much you allow to deviate from your prompt, etc.
So yeah, AI could certainly hit a limit, and very quickly, if you keep settings the same all the time and use the same generic-style of prompting. The same prompt can get a VERY different image from just a minor adjustment, and there's many adjustments you can makeIve seen some very impressive AI generated images. Id say the majority of them are not great.
Probably a lot of people just typing Asian Girl. Would the AI the assume that the majority is the standard and refer to garbage as the standard? Just typing Cat Girl red dress gives 4 different images but similar in some respects.Most of the decent "art" Ive ran across uses a specific theme / art style as a prompt. Something like Steam punk, Giger, street art might produce something interesting. How much control you have over what the AI decides you want is something I question. Do you say wow thats exactly what I imagined and the AI read my mind or do you say Wow, the AI did a great job. Time to take credit for its masterpiece?
Thats because I dont think AI will work well for someone if they are just expecting to type in a prompt and have it shit out a Picasso. It almost never happens. There's a lot more steps to it than most people are being led to believe. Usually what it spits out after a prompt is just step 1. Then it needs to be done again and again with good sections kept and just altering stuff needing fixing or made better (done through better or different prompts, or increasing/decreasing certain settings, etc..) and KNOWING how and what to do (which, yes, does make it "art").
Same as changing/fixing/updating etc in DAZ would be considered art, this is similar just using a different set of tools.In Picasso times, he would look down on DAZ too, and every other digital form of art. I can almost hear him damning it, saying none of it is done using authentic brushes, etc.. times change, the tools change, but the process remains the same.
I follow several people on Tumblr who do AI renders and I've seen them improve month by month. Their individual fetishes are are apparent but the work is getting much, much better. It shows that good images are possible but it takes work and a lot of experimentation.
This persons stuff is bizarre and strange but very well done...
https://www.tumbex.com/andysfantasie.tumblr/posts"Ive seen some very impressive AI generated images. Id say the majority of them are not great."
100% Adam. But not unique to AI. Look through the Daz galleries and the same would apply.
Ok...
Lets say I key in a few prompts.
Lucy Liu in a v-neck red dress biting Wynona Riders neck dressed in a black evening gown. Backround is a full moon on a cloudless stary night. Lets add a few more words like realistic, high quality, photograph ect... whatever paragraph of word salad gets me where i need to be.
The AI kicks out 8 images and one looks near what I want.
I mark that image then edit the prompts like longer dress, black hair, more stars ect.. out of 8 images I pick the one i like and share it on DA.
Now this is hypothetical since ive no clue what I'm doing but I'm assuming this is the very basics on how it works. Im sure there are complex formulas and premade templates.
Am I close or totally wrong?
"Ive seen some very impressive AI generated images. Id say the majority of them are not great."
100% Adam. But not unique to AI. Look through the Daz galleries and the same would apply.
I guess the same 'rules" apply to AI and DS..
In fact it goes for all creative work. The more time you spend on a piece, the more you will be able to pay attention to detail, keep tweaking until it feels right. ( What is NOT a guaranty for a good piece)
You could type a prompt like :
Sexy woman on a beach
AI will provide, but as said above all images are fairly equal.
IF you tell ( what I call) as story to AI, it will be much more precise to cater you desire.
IN DS.
I could drape a sexy woman on a beach, but it would be a boring picture. So you add props, more figures, a nice HDRI, maybe additional (ghost)lights, a bit of wind.
All that could make the pic come to life( again, no guaranty)I used to have a kinda old desktop, so I was limited to boring pics, and soon the fun of creating was gone. Since I got more power to work with, I sometimes take a day or two to make a single pic.
AI is just starting, and I guess the "power" for users will increase ( it will be simpler to create a pic), so it is still playing catch up, but it's getting there fast.
I see AI and DS only as a tool to enable me to create what I have in my mind, just like my preformed stick I use to sculpt clay with. I got a lot of different ones, each with it's own "speciality).
So it all comes down to what tool do you prefer.
- You must be logged in to reply to this topic.
