Midkemia is a pony based model that allows for the creation of anime, realism, semi-realism and several other styles from within the same model. It requires a little bit of prompt tuning for each specific variant but works pretty well.
This model came about as a challenge to myself for members of the UMI-AI discord server but decided to share it when it seemed well liked.
Any lora that work on Pony should be well behaved in Midkemia, to a lesser degree that will apply to SDXL lora as well but your mileage may vary.
Some of the Pixel Art images in the previews were genned with the Palletizer extension for AUTO1111/SDForge and ADetailer was used in most images. However none of the images have been hires fixed.
I hope you'll like the model and if there's any questions or help needed I've linked the discord above!
Description
Initial release to Civitai
FAQ
Comments (17)
I liking it a lot so far, I enjoy models that kill film grain but maintain some form of realism, yours can do that, it also does not have much trouble with colored skins which is good (tested on Sylvanas), but the downside is that as it kills noise it also removes these small textures that 'pop' around the character or cloth, like tiny strands of hair (think outline), if you can fix that it will be a great model, I might replace it with Vendo's or Cyber's eventually after some more testing as I need to test backgrounds, ethnicities and cum.
After many months and trying many models I have a strong belief that pony v6 doesn't play well with realistic stuff, it just doesn't. However, this model comes really close to what it should be, really close. And it doesn't ruin the eyes too much when the face is far, many "realistic" models (which are really closer to 2.5D) produce horrors more often than not. The eye problem is probably why you used ADetailer, and I can say that upscaling fixes this as well (better inpaint and then upscale).
Also this model allows easily control style (anime, cartoon, pixel, etc.) out of the box. It also plays well with loras, both pony and xl.
Lately I used my own very simple merge that is giving somewhat decent 2.5D-reaslistic effect, but now I'm testing this model and I like it a lot.
What did you merge in to get this result?
I really like the results of this checkpoint, but when generating Im coming across some weird issues.
Some promts seem to drive it crazy for some reason. For example I made a girl underwater and it completely destroys the picture (the whole picture completely pixelated) but if I put her "sitting" or "swimming" the image comes fine but if I use "underwater" without adding anything else it pixelates the entire picture. Same if I put her on the beach, but the image cames out well if I put her on a café.
Today with a lora the gen seemed to go fine, then when the after detailer comes in, it pixelates the face of the girl. The rest of the picture works well but the after detailer pixelates the face of the girl.
But I use the same lora, the same gen, but I put the girl in other position or scenario and doesn't give me that issue, it comes out just fine
Is kind of annoying because I dont change the settings between gens. But some particular prompts (nothing in common just random prompts from the ones I try) break the picture entirely.
I've not experienced this myself yet, would you mind giving me the complete prompt so I can evaluate the issue and resolve it? :)
I experimented a bit more with the model and I found what was giving the weird results, It was one of the quality tags that seems to turn crazy the model. So it wass my prompt fault.
As said before, great model
Really nice checkpoint thanks for sharing
This model is so weird. Not necessarily a bad thing, but... It's so unpredictable. I could swear there's so much SDXL/Animagine mixed in the Pony that whether you use some qualifiers from Animagine or Pony, the image style bounces back and forth between the two in a very stark sharp way.
Appreciate the feedback, this is a pretty old version of the model at this point. If you're interested I can upload the latest version here instead of huggingface.
@tinuva For sure! I think I got access on your HF though, I just wonder which file that'd be, haha
@tinuva You have a new one? That'd be great to see it! And on HF which is it, MidkemiaAnimeTV?
@tinuva Okay, I got confused with another repo, lol, I don't actually know your HF link, lol, so yeah, I'd be very interested in seeing your updated version here!
@princess_daphie Sorry for the late reply but here you go, your mileage may vary tho! https://huggingface.co/Tinuva/
@tinuva Thanks! I compared the SHAs of the models on your repo and the one you uploaded here on Civitai is your Experimental V3. I'm testing the others. I don't think they're as wild as the one you posted here though, lol
@princess_daphie I realized afterwards that I never actually uploaded the one I was using last whoops! I'll set it to upload overnight, appreciate you giving them a try even if they're not so good!
@tinuva yeah, i mean, i ended up settling for the one posted here, it really pushes interesting images! if anything, the crazy mood swings could be considered a feature, haha! can't wait to try your new one!
@princess_daphie It's uploading right now, it has a few extra wildcards included with chars, styles and techniques it can definitely do
Can you add the other models from huggingface here to show a preview of how they look? I prefer seeing how they look before making the commitment of downloading. The Artinuva sounds interesting, as do the others.