Stable Diffusion, NovelAI, Machine Learning Art - AI art generation discussion and image dump

  • 🐕 I am attempting to get the site runnning as fast as possible. If you are experiencing slow page load times, please report it.
Post Reply said:
It kills me to have a 6900xt in my Windows machine that I can't use effectively.
Click to expand...
I don't get why radeons suck so much at AI, what gives? isn't the (used) MI25 like the recommended cheap way to so SD? are the gaming cards crippled or what?
 
cybertoaster said:
I don't get why radeons suck so much at AI, what gives? isn't the (used) MI25 like the recommended cheap way to so SD? are the gaming cards crippled or what?
Click to expand...
I've been messing around with running Linux off an external SSD just to see if I can get Stable Diffusion + ROCm running. Limited luck so far. 😞
 
Realistic models are getting better with time. For SD 1.5 there're checkpoints like ICBINP that come pretty close.
 
Problem with realistic pictures is IMO that people go with hyperrealism, things that just look too good and clean to be real, the trick is basically to hide the uncanny valley of SD in picture artifacts, discoloration or imperfect composition, then it can work (better). Examples of what I mean. (not perfect)

somewoman.jpgsomewoman2.jpg


somekindofrobot.jpgsomeflower.jpg

I personally tend to avoid photography though, I always felt it's not the strongest suit of the technology.

I recently discovered ComfyUI and it's an incredibly powerful tool. I like to sometimes generate older realism style oil painting stuff, and while SDXL has the superior image composition, the older models are better at color and detail. I came up with a workflow where I first generate a picture with SDXL, then pass it as latent noise base to an older SD model, but not before making a lineart with controlnet out of it I use as guidance so that the picture doesn't change much at 1024x1024 and the 512x512 SD doesn't go off it's rocker. I had good results with it.

somepainting.jpgsomepainting2.jpg
 
  • Like
  • Winner
  • Informative
Reactions: Neurotypical Mantis, Gog & Magog, A Hot Potato and 6 others
I'm going to have some time off. I'd like to train another LORA on SDXL. Does anyone have any suggestions? I'm still looking more to replicate a style or concept than a character.
To train the style or concept I need the following:
  • The style must have at least 80 images of similar aspect ratios.
  • The images should contain a subject or character with that style of concept represented. That subject needs to be centered and alone.
  • No images which are illegal to procure or a style or subject illegal to reproduce.
  • No scat or really disgusting shit.
You don't need to upload the images just post a link or the name of an artist or a concept.

Edit:Yes Chris has already been done.
 
Last edited:
  • Like
Reactions: A Hot Potato and The Mass Shooter Ron Soye
macrodegenerate said:
I'm going to have some time off. I'd like to train another LORA on SDXL. Does anyone have any suggestions? I'm still looking more to replicate a style or concept than a character.
To train the style or concept I need the following:
  • The style must have at least 80 images of similar aspect ratios.
  • The images should contain a subject or character with that style of concept represented. That subject needs to be centered and alone.
  • No images which are illegal to procure or a style or subject illegal to reproduce.
  • No scat or really disgusting shit.
You don't need to upload the images just post a link or the name of an artist or a concept.

Edit:Yes Chris has already been done.
Click to expand...
Takato Yamamoto
 
  • Informative
  • Like
Reactions: A Hot Potato and The Mass Shooter Ron Soye
Just reporting in, I spent time seriously sitting down and using AI to write. My opinion on AI and writing is now solidified that if you are in a field that requires writing like long form story telling, manuscript writing, and other creative industries of the literary variety and are against its use; you're shooting your own foot. In the span of two weeks, I've been able to use ChatGPT and Sudowrite to write a trilogy of novels. Over 250K words in two business weeks, 450K words if you want to include the character bible, outline, and "lore book". As someone who has carpal tunnel in both arms, I'm not turning back any time soon. If I could find some AI dictation tools, I'd be set for life.

Just out of curiosity I asked GPT to create a blurb based on the chat log and it isn't too far off to what I was reading as the novels were pasted together and re-written. Essentially, I view this workflow as I view a painter's sketch. Lots of potential for human creativity while also having guard rails to keep you on track. Blurb for the first novel Raw from GPT:

Viral Genesis

In the heart of Neon Nexus, where neon lights clash with shadowed alleys, the battle for control rages silently. Helix Dynamics, a corporation at the pinnacle of bioengineering, tightens its grip on the populace through a groundbreaking yet insidious nanovirus, promising evolution but delivering subjugation. Dr. Evelyn Reyes, once their celebrated scientist, becomes their most dangerous adversary after unearthing the truth. Teamed with Alex Kane, the enigmatic leader of Owl Watch, and the formidable Blue72, they take to the streets to rally against the corporate behemoth.

As gang wars flare between the Lumina Knights, GridLockers, and Steel Valkyries, alliances form and fracture in the battle against a common enemy. From the Cathedral Nexus, an ancient monument with mysterious origins, to the Tech Bazaar, the bustling hub of innovation, the rebels must navigate a city of secrets, dangers, and potential betrayals.

But when the line between human and machine blurs, and emotions become programmable, who can you trust? Dive into "Viral Void", where heroes are forged in defiance, and salvation hinges on unity. The resistance has begun. Will Neon Nexus be liberated or lost forever?

All of this started with one sentence: An evil corporation wants to use emotional leverage to influence the community at large, cyberpunk setting, genre: action thriller science fiction. I love the world we live in.
 
  • Like
  • Winner
Reactions: Firewater, Irrational Exuberance, A Hot Potato and 4 others
Post Reply said:
My 2060 12gb is just enough to scrape by, though finding the right settings to get SD XL not to crash was a little tricky. It kills me to have a 6900xt in my Windows machine that I can't use effectively.
Click to expand...
I used a 6800xt on Ubuntu when I was messing around with SD and it worked great.

Where are you getting stuck? The two things that tripped me up were remembering to add the user to render and video with the usermod command and remembering to install PyTorch from the python virtual environment.

There was a guide I might be able to find. I haven’t messed around with it since January.
 
macrodegenerate said:
I'm going to have some time off. I'd like to train another LORA on SDXL. Does anyone have any suggestions? I'm still looking more to replicate a style or concept than a character.
To train the style or concept I need the following:
  • The style must have at least 80 images of similar aspect ratios.
  • The images should contain a subject or character with that style of concept represented. That subject needs to be centered and alone.
  • No images which are illegal to procure or a style or subject illegal to reproduce.
  • No scat or really disgusting shit.
You don't need to upload the images just post a link or the name of an artist or a concept.

Edit:Yes Chris has already been done.
Click to expand...
goli, the main illustrator of the beatmania iidx games. he had two artbooks that can be found on that sad panda website. (or maybe it does already exist a lora of his style?)
 
I might as well dump this image I just made. I was discussing AMD with Stable Diffusion in the Enthusiast thread and did a quick test to let someone know how quick a 7900XT is with AI art generation. This was SDXL model generating at 1,024x1,024. No separate upscaling and with a second refinement pass. It took 34 seconds. For someone who remembers ray-tracing a 3D ball on an Amiga over about 20 minutes, this is amazing.

Behold: A robot kiwi!

robot_kiwi.png
 
  • Like
  • Winner
  • Feels
Reactions: Vash the Stampede, Firewater, plagiarismmaximum and 9 others
macrodegenerate said:
What's your setup? What model are you using and how did you get past the 2k context runaway problem? Seriously impressed.
Click to expand...
It is unfortunately all paid for services that do not have the direct models actively listed; aside from ChatGPT 4. Sudowrite is a web service that has some really great features for authors. 2k Context I didn't really solve it on my local models, I just end up working around it. Although a lot of the text is generated through AI, there is a lot of my own words and editing that went into it. Basically, when I encounter the problem I either not use it in full and regenerate or I take a bit and piece of it.

If I were to take on a virgin process, this is what I'd do:

Ask Chatgpt to list 10-15 story concepts of a particular genre

EX
List me 10 best selling [genre] story pitches.
Click to expand...
List me 10 best selling [genre] story pitches that have compelling characters and stakes.
Click to expand...

When I find one that I like or one that I think would be interesting, I then ask GPT to create a synopsis for a story using the selected Pitch.
When I got a three act synopsis I like, I edit it, bounce it back and forth between editing and GPT until I have something complete. From there I start creating the world building such as themes, settings, characters, religions, gangs, factions, organizations, businesses, infrastructure and sometimes politics. I record all of this into a "Novel Bible" Basically a big brain dump of information that'll probably never be used in the finished story. From there I collaborate with GPT in creating character profiles, Who are they, what do they do, what is their relationship to the novel as a whole, etc. Once I feel I have a good understanding of this world, narrative, themes etc. I move on to outlining through GPT.

Using the Synopsis I ask GPT to create an outline for Act 1 Act 2 and Act 3 separating each into their own novels. That way I can trim down the amount of characters I initially created as well as the settings, themes, etc.

For each act I use a 24 chapter novel outline.
Ex. Prompt GPT
Using the synopsis below, create a detailed outline of the novel fleshing out additional details and
Click to expand...

I essentially just rinse and repeat allowing ChatGPT to do a lot of the heavy lifting. I would let SudoWrite do more but it is capped at a maximum number of words per month so I use text generation on it sparingly. This process is adapted from another author that pretty much taught me everything I needed to know about establishing a basic workflow.

I really do wish that I could have a fully local workflow but until I start making some fuck you money I ain't gonna have the processing power for it all. I'd love to have an entirely autonomous AI Digital "typewriter" laptop to be more mobile. A man can dream right?
 
woowie queen said:
goli, the main illustrator of the beatmania iidx games. he had two artbooks that can be found on that sad panda website. (or maybe it does already exist a lora of his style?)
Click to expand...
Well the art book has 20 images that are similar enough. Plain white backgrounds.
 
You must log in or register to reply here.
Back