Finding all my models or profile elsewhere than Civitai is without my consent !
Please don't repost my files. Send people to my Civitai profile, thanks :)
-
Because I like to do weird things...
Here are Axsys V2 and V3. Both beta versions ONLY FOR TESTING. Don't expect much from those ! To play safe, go for V1. Before publishing V2 and V3, I made many attemps with code errors or more ugly than ever ! These are the least ugly but still are (wow !) so ugly !
ALL VERSIONS NEED VAE.
V3 is a merge from V2 - Illustrious and XL. As a result it will work with Illustrious and XL loras...
But the results are most of the time ugly !
V2 is a merge from V1 - Pony and Illustrious. As a result it will work with illustrious and Pony loras and some XL loras too (because Axsys V1 was already a merge from Pony and XL and was used for this merge).
But the results are most of the time ugly !
V1 PONY :
This is my first ressource ever done anywhere ^^
I was virtualy drunk the other day and decided to make a Flux checkpoint but my computer refused to do it (because of potato hardware) so I ended doing some random merges.
This is Axsys v1, a ramdom Pony merge with an ugly Sdxl base.
This checkpoint is a result of multiple merges, like 7 or 8... the first and second were weird ugly results since i tried to merge Pony with SdXL but then, from an ugly something it became better and better. My merges were totaly random.
The result is a merge that can do many things... Some good and some totaly ugly :)
It can do anime, realistic, realistic mix with anime, nsfw... I haven't tested it in every situation but i'm happy with the results it gave me. It is also not bad or good in img2img.
Try it, I'll give buzz to some of the generations you do with my model.
PARAMETERS
I tried using score_9_up and the others and sometimes not and didn't see any improvement on quality so you do what you want ^^
Sampler : DPM++ 2M SDE / DPM++ SDE and maybe others
Steps : 20 to 50 : i'd say 30-35 is good and produce sharper images than 20.
CFG : I only tried 5 to 7
Resolutions : do as you want, as i don't always use aspect ratios myself ^^
VAE : Normal Sdxl VAE or Normal "Fix16" but others might work but a VAE is needed
I didn't try hi-res or anything else yet.
OTHER
Hands are sometimes ugly and faces too when not near the screen so you can use Adetailer for better results and some negative promps. I didn't try embeddings that improve quality.
NSFW
I know this can be quite a passionate subject and yes you can do this kind of art with this model but it can be better with some loras to help produce what you want..
__________________________________________________
In the future I plan to upgrade my potato pc to do Trained checkpoints but before that, I'll do some other random merges and loras. An XL version will be ready soon :)
Description
This is v1. I plan to do a v2
FAQ
Comments (17)
I think I enjoyed the description more than any of the actual Checkpoints I've downloaded this month. I'll take it for a spin. :) Cheers
So you give buzz? You are something.
About the checkpoint, it seems a normal semireal model sharpen to max. Who like this style download immediately.
Yeah I give buzz :) For nice, it's like 20, very nice 50, top of the line 100 and exeptional... well try it and you'll see ;)
With Stability Matrix's Inference the output images have like some color dulling filter applied - isn't always bad. But what ruins the model is that there are a lot of white spots on almost all of the images. Requires a Refiner to get rid of the issues above.
The training data is unique though.
Keep training.
Thanks for reporting this issue with Stability Matrix's interference :)
I've never had any spots using Forge (in Stability Matrix) but yes, I should maybe try my models in other programs as well.
I've done some testing and that issue with white dots and faded image or filter also happen with some other checkpoints, even some famous ones. If you can, use Automatic 1111 or Forge or even Foocus or any other Stability Matrix has for packages ;)
@Z1os in ComfyUI, I was able to correct this by loading a different VAE. Works great now!
@omegazero983739 Yes with the right VAE, it's all great ;)
@omegazero983739 What vae did you use though if it isn't as secret? Usually models have a baked-in VAE, I'm used to that. But if there's a VAE that fixes white spots I'd like to have it for models without VAE.
@lyagushka420 I think any VAE will do the job for the white dots problem. I usually use "SDXL VAE fix 16" and I haven't had any problems with it, for Pony, Sdxl and some Illustrious. But I have many VAEs so if one doesn't work well, I can change. (some are for anime only)
@Z1os FIX FP16 Errors SDXL VAE solved the issue. I wish knew sooner. I always thought it was the mystical difference between ComfyUI and other frontends.
@lyagushka420 I use Forge so to me ComfyUI is so mysterious :P
@lyagushka420 In comfy, I just added a separate VAE node, it had a few VAEs listed, I just chose the one for SDXL.
But do you guys know how to fix random color artefacts? Not specific to this model. I don't know if it's also related to VAE. Sometimes the image has a detail with sharp color or there's a random color spot on the image, sometimes the color spot malforms the face - is there a fix for it?
@lyagushka420 I never had this
Ok, its better, than i tought. It can be realistic, and not even ultrasharp. I uploaded various sampler and scheduler combos. Decide, which works best.
Thanks :)