So I been playing around with AI and got these results, opinions?
Head stomped:
Literally ripped apart:
I know AI its a touchy subject with the art community right but I still thought it was worth trying. Not imitating any particular artist’s style with these.
lmao
I love them.
You’re getting some great results in my opinion.
It was not immediately obvious to me that this is AI art, which is the first win. At the same time you’re getting some of the major advantages of AI art: flawless perspective, sophisticated use of FOV/picture plane/lens effects, and a 100% homogenous stylistic tone. In short it makes it look professional to a degree that takes decades of hard work for a human artist to achieve.
What you seem to have been able to do, that most AI art fails at, is imbue it with a clear story-telling perspective and a personal touch. That is useful and impressive.
I’m not sure what your “coding hand” has to do with it, unless you wrote the ML parts and trained it on specific art. Is there a scripting interface for more precise tuning of images?
Yeeh… head stomped…
It was just a manner of speak since I’m a programmer but writing promps has a lot in common with code, most people that fail and give up with SD and others is because they think they can just give a natural language description and the AI will magically piece it together, but those pics you see that look truly impressive have some really long prompts full of calls to different LoRAs, and use specific models for each style, that’s why it looks great.
And yeah forget about a personal touch, unless you train it in your own art that is, or mix different artstyles until you get something that may pass as “original”, but its really hard to get what you really want, most of the time its a case of “close enough”.
In other cases…cognito hazard
Link to the ai ur using?
Its stable diffusion, you’ve to install it in your own PC and need a beefy GPU to get it to run, at least 8GB of VRAM.
Thanks! What a neat experiment.
Number 2 is by far my favorite.
Head stomped is cool as well,and the rest are interesting.
Tried making some fluffies on night cafe but they didn’t turn out great.
They are actually great, wtf!
Not bad!
Jeebus-cripes these are gruesome…
You’re definitely good at the prompting with SD, and your output is really good. Nice job~
Very impressive art. Would you mind sharing the prompts you used?
EVEN THE ROBOTS HATE FLUFFIES!!!
I’m slightly horrified by how often these “fluffies” have human feet
it’s a good start, but quite a few of them look like people in costume or perhaps a bit too goat-like
None of these are proportioned even close to what a person actually making art would imagine a Fluffy Pony to resemble, down to the hips and leg lengths. Some even have the feet of human kids. In clothing.
You didn’t “create” fluffy pony abuse, you used a computer to hodgepodge bad animal/child abuse fanart (which should be a rule violation) from existing pictures. You didn’t “code” anything, or make anything worth sharing.
What a clueless set of statements.
Doesn’t look like a fluffy pony, though the second is more horse-like. Overall they look more like a person / catboy wearing a fluffy jacket. I wouldn’t be surprised if that’s how it interpreted “fluffy pony”.
I recommend using something to swap or interpolate prompts so that you can start with a prompt relating to horses or ponies or mini ponies etc to set the initial composition, and then switch the prompt to include thick or curly fur.
Might be easier to generate a small horse at first and then modify it via img2img.
Could try using GitHub - klimaleksus/stable-diffusion-webui-embedding-merge: Extension for AUTOMATIC1111/stable-diffusion-webui for creating and merging Textual Inversion embeddings at runtime from string literals. to combine tokens within your prompt to create novel concepts like <'pony'+'pig'=/2> or <'horse'+'pig'*0.6=/1.6> which might help get a better base.
Immediately looked AI generated and has mismatched limbs and unrelated concepts (fluffy jacket and anime faces and human limbs). The jacket should be solvable by putting it in the negative prompt after 0.4 or so of the steps have completed, like [jacket:0.4] in the Automatic1111 webui formatting.