CAN'T BE!
This is a lora trained on a large combination of sprites from Pointless, Pointless SOTW, and Pointless Infinity Unfolds. This model boasts a considerable upgrade compared to the last model which was trained on an oddly specific version of fluffyrock (and accidentally at the wrong resolution oops :p). This was trained with a total of 227 sprites on the fantastic Pony Diffusion model at 1024x1024 for all the images (the last model only had 176 sprites). It was trained on only battler sprites, so it's really good at generating humans, but also pretty much anything else PDXL can generate, it understands the style pretty well! I trimmed out the sprites of shadow monsters, some joy mutants, some animals, and repeats characters from the dataset. It was trained on e621-style tags.
NOTES:
-You must generate with the Pony Diffusion model, and ideally, you should generate at 1024x1024 resolution. Technically it can do other aspect ratios, but every image was trained at a static 1024x1024 1:1 aspect ratio (square).
-I specifically tagged infinity franchise members as "jerseyhead", so use that if you want a character wearing that funny 88 jersey.
-You can actually have detailed backgrounds and landscapes, however it was not trained with any "LISA" Parallaxes, so keep that in mind.
-It can do concepts that weren't trained, such as women, or vehicles or what-not, so try it out in those regards. In fact, I guess this could be a 'LISA-styled' lora, since it can do landscapes and backgrounds.
-It wasn't trained on any pretty much any joy mutants, however it can still do monsters and odd creatures pretty well if you prompt it good enough.
-See final example photo for a brief tutorial on how to downscale the generations down to workable sprites using photoshop, if anyone is mad enough to actually make something with this and put it into a project. When upscaling the sprites for the training, I had upscaled them all by 700%, and put them onto a purple background, so it's a guide on how to bring it back down to 1x1 sprite standards.
CREDITS:
Austin Jorgensen - Creator of LISA
Edvinas Kandrotas - Creator of Pointless
All the devs on SOTW - They made most of these sprites!
All the devs on Infinity Unfolds - They made 20-30% of these sprites!
MrFood - He made this model, I'm just uploading it in his place again :P
Example prompt: (See example images, all images should still have their metadata. Just drag the images onto the PNG info tab and import it.)
MU never ever... ToT
Description
FAQ
Comments (15)
Big fan of your 1.5 uploads. Any chance you plan on bringing any of them over to Pony? lol
Yeah it's been a bit, huh? I should probably get around to updating some of those to pony since pony seems like it's gonna stay for some time. Any models in mind you would want updated? I still have all the datasets I trained those models on archived, I'd just have to take the time to re-tag + re-train, and perhaps add any new data to them available.
@ZoochMan Pony seems to know Haydee pretty well. It knows Tali but loses a lot of details on her suit. What I'd really like to see are the Artificer and Huntress ones brought over. You're the only person I ever saw giving Risk of Rain any love. (Its probably asking too much, but the other RoR characters would be pretty rad).
@sliptrick Yeah I've spent the last 20ish minutes just trying out what Pony can do natively that I've trained in the past. That's kind of a big reason I made all my models when I did, there wasn't any other way to generate them.
But now, these models are getting smart enough to do things natively, which is awesome.
I'll see about the Artificer and Huntress ones, but no promises. Any other characters you specifically want? I still need to see what models are even worthwhile of updating...
@ZoochMan I like the military ones like the combine ones and the haven trooper outfit. They're good for using at middle weights and getting actually good tacticool outfits.
I haven't tried sangheili in pony, but I wouldn't be surprised if it knows it. That and I think your spartan lora always worked better than other ones I tried. Like, it works better with style loras.
@ZoochMan If you meant other RoR characters (lol) Mercenary and Bandit would be cool.
any chance you could share the traning process? I want to make some pixel art lora myself, but I've found it very challenging
Pixel art is inherently very hard for Stable Diffusion to do, since it doesn't generate at small resolutions, . So for me to make a model like this I had to upscale each sprite by 700% until it was nice and big for a 1024x1024 background. After that, I used photoshop generously to automate each of the transparent sprites onto a purple background, and made sure to center them in the middle bottom. After that, I just trained the model as expected, however I think I had to train this model for MUCH longer than a regular one, I had to run this one for 10,000 steps for it to understand the style. I had numerous trainings fail since they were much lower than this...
It helps that the LISA games typically have all their 'battle' sprites in a single folder in any of the games where every single 'battle' sprite are in nice, uniform files. It made datasetting easy.
In short: Find a good dataset, isolate the images if they aren't already, use photoshop automation to upscale + add a background (recommended), tag them with 'Z3D-E621-Convnext', then train
@ZoochMan Thank you for the upscaling tip, I was seriously pasting a lot of stuff in a 1024x square, but it wasn't working
@Daru_22 It helps if you got a lot of sprites. You just gotta make them as consistent as possible, and even then it'll never be perfect. Seriously though, don't underestimate automating stuff in photoshop, it takes a moment to learn but it can turn hours of work into minutes (I can't imagine manually pasting all of these sprites onto a purple background and manually saving each one)
Never give up!
An uneducated question: is there a prompt you have tested, which works to make the sprites consistant? So I can include my prompt to, walk, face left right, face up down etc?
This model is fairly inconsistent, so no, not particularly. I would recommend working from one of the example images I posted, and experiment from there.
As for facing different directions... This was trained solely on "battle" sprites, meaning about 90% of the characters are just facing the viewer. If you (emphasize) the right tags enough, you might be able to achieve results with certain tags.
@ZoochMan thanks, and keep up the great work. We all appreciate it
@Ooti Of course. I'll eventually retrain this since this is my pet project I keep coming back to. With LISA: The Undone released, maybe I'll incorporate that, and up the steps trained. I'm dedicated to making a decent sprite generator against the odds
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.