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

No. Custom blend of various models. The custom blend stuff shows when I proompt for anime shit since it makes it all look hyper realistic. I have another that does the exact opposite to where all realistic stuff look like anime or at least 3DCG.
What's the mix, nigger?
Mine always end up shitty... I really only want the celebrity tag recognition to work with NAI.

New Stable Diffusion update reportedly nerfs ability to use celebrity likenesses, create NSFW images, and copy artist styles.

Archive

I think basically all this means is more third-party forks are coming because the tiddy genie isn't going back in the bottle.

Anyway here's one of the first things I generated after getting this shit installed last night: a demon playing a guitar solo in the style of Frank Frazetta

View attachment 3962088
Well, they're fucking with it to intentionally avoid something. This will inevitably make it useless. There's always collateral damage. Also the utility is down a ton if you can't generate celebrities.
 
  • Thunk-Provoking
Reactions: SomeBirds
Well, they're fucking with it to intentionally avoid something. This will inevitably make it useless. There's always collateral damage. Also the utility is down a ton if you can't generate celebrities.
They're trying to avoid legal issues. The guy in charge said he's worried about governments framing AI as a CP and deepfake generator and using that as an excuse to ban everything.

If it's any consolation, he also said the new model should be much easier to further train/modify.
 
I think basically all this means is more third-party forks are coming because the tiddy genie isn't going back in the bottle.
New version of WaifuDiffusion drops tomorrow iirc and according to the authors is supposed to supersede everything "to the point where you don't need to merge".

...They said that before 2.0 dropped (it's based on 1.5), but it does have a bunch of improvements, higher res training (though 640x as opposed to 2.0's 768x), and better anatomy etc. I wouldn't be surprised if it is a substantial improvement considering base S-D has never been that great for boobies and the cafe set test release (which it integrates) was promising.
 
The process of using multiple re-images to get a desired outcome can be interesting. Here's how I turned Chud into Chud HD. Click any thumbnail to see higher res and more detail.

1. Start with Chud. I couldn't get good results with this. It didn't want to change it from a cartoon. Everything it generated was a cartoon of some sort. I had to force the issue.
OtsYCXCaltlEeLzVZXNzJzmO.png

2. I took it into Photoshop and thinned down/ removed some lines and added some simple shading. This should force the AI into seeing it as a 3d object.
chud1.jpg

3. First successful result! It made it 3d, but looks like an unrealistic 3D model. The glasses also look strange. This had a hard time giving me good results as well. I decided to use a separate AI upscaler to see if using higher res image would give me better results. I also google image searched up some sunglasses and photoshopped them over the weird goggles, and morphed his mouth to match the original better.
16a2fbf083d2430baf633f5318778fa6.jpg

4. Result of AI upscale + new glasses + changed mouth. The upscaler not only made it higher res, but the AI in the upscaler recognized it as a face and added a more realistic texture to the skin when interpreting the detail it needed to add. Great bonus. This should only need one run through Stable diffusion as a low prompt strength setting to get it to look photorealistic.
efb328295050446a86e392a8eb8ae6b6.jpg

5. Final result. Chud HD is born! More realistic overall, with wrinkles and creases as well the subtle discolorations that skin normally has. It also made the hair look more natural and added stubble.
chud HD.jpg

OtsYCXCaltlEeLzVZXNzJzmO.pngtochud HD.jpg
 
Introduction

I was going to post this awhile ago to help people starting out on Stable Diffusion, but the farms were under fire. I have spent hours of time generating random shit, this software is very addictive. I highly advise limiting yourself to using on occasion. Unless you need references, as a perfectionist, I need the absolute best and when the program decides to give your character four extra limbs...lets just say you waste a lot of time tweaking prompts. Nonetheless, it still generates really great work! As I will display my favorite results with prompts, models, clipping, and sampling methods I have used in these batches.

Bare Bones Starter Advice For Text2image

1. Be careful of what model checkpoints you are downloading, lets not pretend people do not get boned downloading stuff from the Inter Webs. You are better off getting models from hugging face 2.0 or training your own hyper network to make your preferred models tailored to your liking. Hugging face 2.0 has implemented a feature to test if a model is safe to download. All models I have used are safe and available to the public.

2. For better generation outcomes, change clipping layers to 2 as opposed to 1.

3. Every sampling method is different, meaning they require a certain amount of steps to come out clean. This does not mean you are not going to get squirrely results. I found using Euler to test my prompts first is the quickest method to discovering what works and does not. If I enjoy the outcome result, I will use DPM 2 A.


EULER - 30 steps EULER A - 10-100 steps HEUN - 10-30 steps LMS - 10-50 steps PLMS - 30-60 steps DDIM - 20-60 steps DPM - 10 steps DPM FAST - 50-150 steps DPM 2 - 10-40 steps DPM 2 A - 40-100 steps DPM 2 A KARRAS - 40-150 steps DPM 2 KARRAS - 30 steps

4. Each Model has been trained differently. If you are getting something completely different than expected check your checkpoints. Here are the models I have used in this post.
Nova A.I. [925997e9], Stable Diffusion 1.4 [7460a6fa], Stable Diffusion 1.5 pruned [a9263745], Waifu Diffusion V1-3 Float 16 [84692140], and Anything-V3-prunned [2700c435]

5. Most importantly never give up! Do not be shy! The Internet is a great tool to troubleshoot problems with this software! We are here to help! From my personal experience, when first downloading the software, anything I generated came out pure black. No matter what I did, until I learned I had to tweak the webui user batch file limiting RAM usage. I was going to quit, I said to myself constantly "No I will figure this out and make it work" I eventually pulled through. This is my personal favorite youtuber for Stable Diffusion advice or recommendations- https://www.youtube.com/@Aitrepreneur Highly recommend!


-My Favorite Prompt Results-

Juke-Box Mechs - masterpiece, best quality, masterpiece, Richard Corben and Jim Burns art style, jukebox mecha, full body
Negative prompt: lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, artist name
Steps: 50, Sampler: DPM2 a, CFG scale: 12, Seed: 1940982481, Size: 512x512, Model hash: 925997e9
00057-1940982481-masterpiece, best quality, masterpiece, Richard Corben and Jim Burns art styl...png

Seed: 2470457944
00067-2470457944-masterpiece, best quality, masterpiece, Richard Corben and Jim Burns art styl...png

Seed: 1940982485
00061-1940982485-masterpiece, best quality, masterpiece, Richard Corben and Jim Burns art styl...png
(Context- Who doesn't love giant fighting robots?)

Nazi Germany Mechs War Torn Landscape
- masterpiece, best quality, masterpiece, Yoshiyuki Tomino, Code Geass art style, Nazi Germany Mecha, realistic, war torn landscape
Negative prompt: lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, artist name, blurry face
Steps: 80, Sampler: Euler a, CFG scale: 12, Seed: 754418164, Size: 512x512, Model hash: 84692140
00297-754418164-masterpiece, best quality, masterpiece, Yoshiyuki Tomino, Code Geass art style...png

Seed: 2752675448
00295-2752675448-masterpiece, best quality, masterpiece, Yoshiyuki Tomino, Code Geass art styl...png

Seed: 754418167
00300-754418167-masterpiece, best quality, masterpiece, Yoshiyuki Tomino, Code Geass art style...png
(Context- Who doesn't love giant German fighting robots?)

Oil Painting of a Beautiful Blonde Hair Blue Eyed Female Vampire
- Beautiful blonde hair blue eyed female vampire, (full body), horror, dark setting, FEMININE,((PERFECT FACE)),((SEXY FACE)),((DETAILED PUPILS)),(greg rutkowski) ,trending on artstation, artgerm,(Simon Bisley),OIL PAINTING, (((SMALL BREAST)),((TONED ABS)),EVOCATIVE POSE, SMIRK,LOOK AT VIEWER, ((black dress)),(INTRICATE),(HIGH DETAIL),SHARP
Negative prompt: ((nipple)), ((((ugly)))), (((duplicate))), ((morbid)), ((mutilated)), [out of frame], extra fingers, mutated hands, ((poorly drawn hands)), ((poorly drawn face)), (((mutation))), (((deformed))), ((ugly)), blurry, ((bad anatomy)), (((bad proportions))), ((extra limbs)), cloned face, (((disfigured))). (((more than 2 nipples))). out of frame, ugly, extra limbs, (bad anatomy), gross proportions, (malformed limbs), ((missing arms)), ((missing legs)), (((extra arms))), (((extra legs))), mutated hands, (fused fingers), (too many fingers), (((long neck)))
Steps: 40, Sampler: Euler, CFG scale: 20, Seed: 3562177528, Size: 512x768, Model hash: a9263745
00147-3562177528-Beautiful blonde hair blue eyed female vampire, (full body), horror, dark set...png
(Context- Found the vampire prompt on a Stable Diffusion subreddit. It generates very good results for portraits to hang up in your house! I made quite a few changes, but ultimately got tired of the prompt really quick.)

Terrifying Phantom
- masterpiece, best quality, masterpiece, (upper body portrait), art by Mike Manley, art by Richard Corben, art by Jack Kirby, a terrifying phantom, trending on artstation, highly detailed
Negative prompt: (multiple characters, lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, artist name, (poorly drawn hands), (poorly drawn face), blurry, (bad anatomy), (bad proportions), (extra limbs)
Steps: 40, Sampler: Euler, CFG scale: 12, Seed: 3037420633, Size: 896x1152, Model hash: a9263745
00153-3037420633-masterpiece, best quality, masterpiece, (upper body portrait), art by Mike Ma...png
(Context - This is not what I requested, however it is very neat. I feel like I am witnessing some hellish ghouls escape the depths of hell.)


Iron Phantom
- masterpiece, best quality, masterpiece, (upper body portrait), art by Jack Kirby, art by Shotaro Ishinomori, ((iron phantom)), anti-hero, trending on artstation, highly detailed
Negative prompt: (multiple characters, lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, artist name, (poorly drawn hands), (poorly drawn face), blurry, (bad anatomy), (bad proportions), (extra limbs)
Steps: 30, Sampler: DPM2 a, CFG scale: 13, Seed: 1455667768, Size: 512x512, Model hash: 925997e9, Batch size: 3
00179-1455667768-masterpiece, best quality, masterpiece, (upper body portrait), art by Jack Ki...png

Seed: 1455667770
00181-1455667770-masterpiece, best quality, masterpiece, (upper body portrait), art by Jack Ki...png

Seed: 1455667765
00176-1455667765-masterpiece, best quality, masterpiece, (upper body portrait), art by Jack Ki...png
Note- These will come out better with an increase amount of steps. Either way, I was absolutely astonished of how this prompt came out! It provided me with so many character design references I would have never thought of. Very inspiring. So if you are interested in creating a hero, villain, or anti-hero. Use this prompt!

New Planet in Space - Masterpiece, Best Quality, Masterpiece, new planet in space, background
Steps: 30, Sampler: Euler a, CFG scale: 12, Seed: 3068543151, Size: 1024x1024, Model hash: 7460a6fa
00464-3068543151-Masterpiece, Best Quality, Masterpiece, new planet in space, background.png

Seed: 385226817
00469-385226817-Masterpiece, Best Quality, Masterpiece, new planet in space, background.png

Seed: 385226813
00465-385226813-Masterpiece, Best Quality, Masterpiece, new planet in space, background.png

New Planet in Space Ver 2 - Masterpiece, Best Quality, Masterpiece, new planet in space, background, Carl Heinrich Bloch, Jean-Antoine Watteau, Felice Casorati, Anni Albers
Steps: 30, Sampler: Euler, CFG scale: 12, Seed: 1306629244, Size: 1024x1024, Model hash: 925997e9
00523-1306629244-Masterpiece, Best Quality, Masterpiece, new planet in space, background, Carl...png

Seed: 3010601334
00524-3010601334-Masterpiece, Best Quality, Masterpiece, new planet in space, background, Carl...png

Masterpiece, Best Quality, Masterpiece, new planet in space, background, Carl Heinrich Bloch, Jean-Antoine Watteau, Felice Casorati, Anni Albers
Steps: 40, Sampler: Euler a, CFG scale: 15, Seed: 3010601334, Size: 1024x1024, Model hash: 2700c435
00527-3010601334-Masterpiece, Best Quality, Masterpiece, new planet in space, background, Carl...png
(Context- Using the Nova A.I. model will produce results with characters rather than backgrounds landscapes. It can still create very epic backgrounds but a character will be in the way. The same goes with the new model on the block Anything V3. While these were not the attended results I was looking for, I was still amazed at these unique pieces in particular.)


I hope this inspires great pieces of Art! Happy generating friends and share your prompts!
 
Clip Studio Paint (popular digital drawing program) is adding an experimental AI feature in its next update.

Unsurprisingly, there is backlash.
"NOOOOOOOO STOP MAKING TOOLS THAT'LL MAKE WORK EASIER :mad:"
No doubt Japanese artists are seething too since Clip Studio is much more popular over there. They can whine all they want but there's no doubt that mangakas will sneakily start using it to speed up their workflow and lessen the need for assistants. Anyone who's smart will do the same. Work smarter not harder.
 
There are a couple of people talking about how the AIs in circulation have all been lobotomized when it comes to certain things like generating celebrities, so instead of quoting a bunch of people I'm just going to generally address the thread.

One of the first things we tried to ask the AIs for was some celebrity stuff, so we quickly realized that the AIs had been crippled when it came to doing certain things. This is what lead us to looking into training. I've left all the links below for anyone else who is interested. It's definitely the fastest way to add famous people to your library.

This is the github page for the WebUI Stable Diffusion package. The DreamBooth extension is installed either from the command line or just from the extensions tab in WebUI. Link

This link is to the Wiki for WebUI Stable Diffusion and is a must read for any user of WebUI. Link

These links are training links. The Benny Cheung links are probably the most useful, but everything in here is good.
Link
Link 2
Link 3
Benny Cheung Link
Benny Cheung Link 2

Unstable Diffusion -- This is a fork of Stable Diffusion where they are making a version that isn't crippled and can generate anything.

Here's some examples:
fyOaqT2.jpg
Mads Mikkelson

jf4Q1Cu.jpgd0VYwyI.jpg
Charlie Cox from the Daredevil TV series

For some weird reason we could never get good results for fantasy creatures, so here's an example of after we trained it to recognize centaurs. This still needs work; we have another training round planned for later.
centaur2.png

Unfortunately the next test lined up is whether training massive image batches solves some of the loss issues when merging a DreamBooth trained library, so no more test results until we can finish processing 1000 images or so and then running them (which will probably require a few days straight of training).
 
Anime, China and art powerleveling autism here, but QQ has an AI Art generator, and I decided to test out its potential with some of my art. Turns out it’s also free as well, so people can mess around with it for a long time.
I tested one for 6 times, 3 times and 2 times. I am surprised that it does it fine.
74ED7976-1C65-4899-A0A8-F9EBDEFFCB03.jpeg5391D51F-C92D-427B-B48E-B4C4A099FC60.jpegEB08CFAB-5426-4AF0-827A-C857DDA759BA.jpeg9BCF13A3-7AB2-44F0-8809-98CDAB1C3414.jpeg14D11CCB-7441-49B8-8499-23C2CB7DB5B2.jpeg58D7E601-4EFF-42F2-B9D5-CC3695F3B1AC.jpeg
617B4B4A-BAD7-4EF2-81C3-F7ECA6501D59.jpegCFA18E4B-8CB1-4D0D-B2AF-FCE984D07901.jpeg10878E1F-2FDA-41BD-A3A4-698FA3728E82.jpeg
ACDF54B7-CA8D-45FA-BAB6-F776E94BB2C6.jpegECB37B00-2BF3-49F2-8EA1-88807EB3D3E9.jpeg39DB0CAD-2E36-4F9C-B8A9-E12DA9045263.jpeg
4386EDDE-EF41-40B4-8EAE-C8EC19C04B96.jpegD2768308-AE31-42BB-B9FC-180F7CE98F1E.jpeg
4EEA5A47-D54C-4EB0-821F-82EB1271A7DF.jpeg26C74144-487F-4641-B64B-65F442000D64.jpeg
Other than the annoying promotion interface thingy for the generated picture, it does most anime-styled art fine. But still, AI art is still emotionless and not as lively as the other way around.
 
The process of using multiple re-images to get a desired outcome can be interesting. Here's how I turned Chud into Chud HD. Click any thumbnail to see higher res and more detail.

1. Start with Chud. I couldn't get good results with this. It didn't want to change it from a cartoon. Everything it generated was a cartoon of some sort. I had to force the issue.
View attachment 3970293

2. I took it into Photoshop and thinned down/ removed some lines and added some simple shading. This should force the AI into seeing it as a 3d object.
View attachment 3970302

3. First successful result! It made it 3d, but looks like an unrealistic 3D model. The glasses also look strange. This had a hard time giving me good results as well. I decided to use a separate AI upscaler to see if using higher res image would give me better results. I also google image searched up some sunglasses and photoshopped them over the weird goggles, and morphed his mouth to match the original better.
View attachment 3970311

4. Result of AI upscale + new glasses + changed mouth. The upscaler not only made it higher res, but the AI in the upscaler recognized it as a face and added a more realistic texture to the skin when interpreting the detail it needed to add. Great bonus. This should only need one run through Stable diffusion as a low prompt strength setting to get it to look photorealistic.
View attachment 3970353

5. Final result. Chud HD is born! More realistic overall, with wrinkles and creases as well the subtle discolorations that skin normally has. It also made the hair look more natural and added stubble.
View attachment 3970380

View attachment 3970293toView attachment 3970380
This looks super good!
Out of curiosity, which upscaler did you use?
 
"NOOOOOOOO STOP MAKING TOOLS THAT'LL MAKE WORK EASIER :mad:"
No doubt Japanese artists are seething too since Clip Studio is much more popular over there. They can whine all they want but there's no doubt that mangakas will sneakily start using it to speed up their workflow and lessen the need for assistants. Anyone who's smart will do the same. Work smarter not harder.
I can easily see it it being used for mundanity like backgrounds. I already know of some artists that either trace over or edit 3D models of environments in order to keep them consistent and save time.
 
  • Agree
Reactions: Ether Being
Back