Fan Art Something Unlimited - Sprite Re-Mix and Fanart Thread

5.00 star(s) 2 Votes

Jimwalrus

Active Member
Sep 15, 2021
912
3,360
There's an model out there, so I looked at what it might produce.
Going by the comments on Civitai it seems that it was trained, in part, on SR7's artwork. So it's just generating variations on that, with perhaps a little added chonk*. Still nice though.
00017-986765277.png 00018-4032607741.png 00020-1810545500.png 00022-1810545502.png 00024-27877257.png 00025-27877258.png 00027-27877260.png

*For the SD nerds on here, this is a perfect example of what happens when both the LoRA and the Checkpoint are trained on much the same image sets. It doesn't (contrary to the various law suits going around) "pull out" the training images from its magical 'data banks'. It just makes reverse engineering something close to those images very easy with the appropriate prompts i.e. it so closely understands the artist's style and the character's design that it's easy to re-generate almost the exact images.
 

DarkDank

Well-Known Member
Oct 26, 2019
1,149
1,550
There are two models of her out there, so I had a go. The results were poor.
I couldn't reliably generate her with larger breasts (she's canonically fourteen; I needed to make her look like Lex was capturing her at 18 for obvious reasons) and still generate the ladybug costume properly. I won't be posting any of the attempts as all were aborted mid-generation and look awful.
Plus the face never really came through as clearly SU-style, it always retained the look / texture of her face on the show.

Edited to add a re-run of the best of a bad bunch:
View attachment 2614633
Thanks for giving it a go! You were under no obligation to so I appreciate it!
It's not bad at all! And I definitly think an "official" Ladybug would look very similar!
 
  • Like
Reactions: Jimwalrus

Jimwalrus

Active Member
Sep 15, 2021
912
3,360
maybe all these ai-cg sr7 crossover pics can move to their own thread...
You may well be right, I certainly feel like I'm monopolising this thread. It does still class as 'fan art' though, IMO.
I am going to have to give it a rest for a while anyway - I've just been rubbered with a load of work to do. :(
 

DarkDank

Well-Known Member
Oct 26, 2019
1,149
1,550
You may well be right, I certainly feel like I'm monopolising this thread. It does still class as 'fan art' though, IMO.
I am going to have to give it a rest for a while anyway - I've just been rubbered with a load of work to do. :(
I think you should keep posting here.
Thread is slowly dying otherwise I think.
Smokeshank of GG (who does....?) wants to keep all the fanart on their Discord and has lured some of the artists here away to that Discord for the "official" fan art channel (after they vet and approve you...for fan art...).

If others want to keep it separate too then majority wins, but my vote is to keep AI in this thread.
 
  • Like
Reactions: Stagger123

Jimwalrus

Active Member
Sep 15, 2021
912
3,360
Personally, I would like to agree that the AI art should move to its own thread, and I'm glad someone else suggested it. I wanted to, but wasn't sure how to say it without coming across as rude or whiny.
There is a general AI thread for AI images, but I think these are a bit too specific for that.

The deluge of my odd AI-generated SU hypotheticals has largely dried up now - I've got work to do, the novelty's worn off for me and I'm also aware it's drowning out actual drawings / sprite-bashes.

I'm open to DMs for commissions, within the usual rules of this forum of course, no guarantee I'll do them though - depends if it tickles my fancy.
 
Last edited:

1138op

Member
Jun 2, 2018
119
535
With all the AI art, all that remains is the invention of reasonable priced hard-light hologram emiters and porn as we know it, or even prostitutes will be a thing of the past.
 
  • Haha
Reactions: Jimwalrus

Jimwalrus

Active Member
Sep 15, 2021
912
3,360
With all the AI art, all that remains is the invention of reasonable priced hard-light hologram emiters and porn as we know it, or even prostitutes will be a thing of the past.
Yeah, but I'd stay away from getting a handjob...
 

Sinner87

Member
Jun 22, 2017
196
2,590
There is a general AI thread for AI images, but I think these are a bit too specific for that.

The deluge of my odd AI-generated SU hypotheticals has largely dried up now - I've got work to do, the novelty's worn off for me and I'm also aware it's drowning out actual drawings / sprite-bashes.

I'm open to DMs for commissions, within the usual rules of this forum of course, no guarantee I'll do them though - depends if it tickles my fancy.
I thinks it's perfectly OK that you post your AI stuff here. It's still fan art and a lot of the stuff you post is really good and is pretty damn close to the games art style. Hell, all I do I sprite bash Sunsetriders7 art. I don't do any real drawing myself.
 

Mr-Fox

Well-Known Member
Jan 24, 2020
1,401
3,793
Some better ones of Tracer - still struggling a little with the goggles either not coming out properly or making the face look a little off. A tip from Mr-Fox in the Stable Diffusion thread with regard to using higher numbers of upscaling steps to resolve issues with LoRAs and it gets quite a bit better though.

View attachment 2614590
View attachment 2614575 View attachment 2614584 View attachment 2614592 View attachment 2614593 View attachment 2614595 View attachment 2614607
It's not quite what I meant. The original tipster is devilkkw. When using hires fix, use 2x sample steps for hires steps, meaning if you use 20 sample steps use 40 hires steps. This is to keep the composition from the sample stage to the hires stage.
That's all. About lora's though. The guide on renty about training lora's is saying that a well trained lora should work well at weight strength 1.0. This is a bit idealistic in my experience. For me it worked best 0.4-0.8 .
The one main "takeaway" from the guide on renty is that slower training is better than fast.
Meaning slower with longer training. The guide talks about things that dampens the training "speed".
Wich checkpoint you are training your lora on is very important. I tried a bunch of different ones and landed on elegance.
I didn't try a SD1.5 basemodel, supposedly it's not great for consistency with different poses.
It all depends of course on the concept of the lora. If you are having problems with getting a lora to work properly despite testing with different checkpoints and trying different weights and work a lot on the prompt, then the Lora needs more work.
Before making that conclusion though, do a lot of testing with different checkpoints and put enough work into the lora and all the possible settings in SD.
The issue with the goggles could be due to the denoising strength, I would also do tests with different cfg scale.

 
  • Like
Reactions: Jimwalrus

Mr-Fox

Well-Known Member
Jan 24, 2020
1,401
3,793
Some better ones of Tracer - still struggling a little with the goggles either not coming out properly or making the face look a little off. A tip from Mr-Fox in the Stable Diffusion thread with regard to using higher numbers of upscaling steps to resolve issues with LoRAs and it gets quite a bit better though.

View attachment 2614590
View attachment 2614575 View attachment 2614584 View attachment 2614592 View attachment 2614593 View attachment 2614595 View attachment 2614607
Looking at the generation data, the settings is a bit out imo. Try 20-30 sample steps with 40-60 hires steps, start high and then reduce a bit at the time, and 8-12 at least cfg scale. Denoising strength you can try 0.24-0.34 or so. Also take a good look at the prompt itself and try different lora weight, 0.4-0.8 . The upscaler is also very important so try a different one. For smooth details try 4x-ultrasharp, for more intricate and realistic details try NMKD Superscale.
 
Last edited:
  • Like
Reactions: Jimwalrus

OhWee

Forum Fanatic
Modder
Game Developer
Jun 17, 2017
5,718
28,976
R.E. this thread and AI generated art.

AS the one that started this thread, I'm good with the AI art being shared here, particularly as long as it relates to SU and/or SR7's art style. Note that SR7 has done Marvel, etc. art as well as DC art so it's the 'art style' that is what is important here.

I mean, if you are emulating SR7, the only other place you'd theoretically post would be in the SR7 asset packs thread, except that said assets thread is for 'official' SR7 artwork, not fanart...

So monopolize away! I think it's not so much that others aren't posting as much, a number of them are still posting at the same frequency actually as before, it's just that the AI generated art has a 'faster' process to generate said art, hence why that content is showing up more often lately.

My personal 'wish' is people putting together new sets of sprites that look like they belong in SU and that can be used to tell a story (so there would need to be sexy fun time sprites as well as just sexy sprites) but apparently the AI is having difficulty with using the same outfit or something, so hopefully that can be overcome...

Also, it'd be nice if more people could help with coding the girlpacks for SU ReImagined, but yeah...
 

OhWee

Forum Fanatic
Modder
Game Developer
Jun 17, 2017
5,718
28,976
Jimwalrus
I got some suggestions in mind for the AI:
  • Hayasee Nagatoro from Please Don't Bully Me, Miss Nagatoro.
  • Kekko Kamen from Kekko Kamen
  • Panthra from Teen Titans
I was thinking that Panthra was in Thundercats, but that's Panthro, a dude...
:ROFLMAO:

There's Cheetara of course!
 
  • Haha
Reactions: DarkDank
5.00 star(s) 2 Votes