A big thanks to civit77899 and AFD_0 for this V5 model and all models going forward
Description
Fixed an issue I was having with the VAE in all my models and this is the result
FAQ
Comments (6)
Sorry, I had to add more of these images in the silk PJ's. The ass was to nice not to :)
The image i did from Tarendax in the silk PJ'S, no matter how hard i tried i couldn't get her to show her ass unless I re-wrote his prompts but I loved his direction, scene he created and I couldn't ruin that. He does some good work and I think very creative, it shows in his images
@jrrytng467 Test_ v5.0 loads just fine! Unfortunately, it's only giving me completely washed-out images with white dots everywhere, so I'm assuming there isn't a VAE baked-in at all for this one??
I'm trying it again while loading an external VAE (standard sdxl_vae) and will report back with what happens. Fingers crossed!
Update; Loading an external VAE fixed the problem! Should look into the process needed to fully bake-in the standard SDXL VAE before posting the final merge. Anyway, first results are looking great! Though the first image had a hand with only 3 fingers and 1 thumb, while the second image had 6 fingers and no thumb, lol. Otherwise, everything is looking really good so far!
Yes, hands I believe are an issue with this model. Im posting one last one for awhile tonight. I hate to speak to soon but I think this one is going to be the best by far. I ran a bunch of tests and looks like I got the hands back under control. Funny issue about the VAE because I got a standard SDXL VAE from this site and is installed in my stable diffusion to load automatically.
Anyways, this new model Im not going to bother putting to much effort into explaining it but Ill tell you real quick. I have 2 creators (you can probably guess) wildzzz, and ironically_irascible. I took what I feel are the 2 best models each that they have. and merged all 4 together 2+2 at 0.75 the nice thing about the 2 creators work is one is strong NSFW and the other is strong SFW <- the base part of the merge and the find a nice balance. They both follow prompts well and both put out beautiful images. Anyways, after this one Im going to learn some new stuff just for my anime model I play around with time to time
@jrrytng467 I'll give that one a try this weekend! The strangest thing is, I've had pretty decent luck with ironically_irascible's models, but none at all using wildzzz's (even though theirs and everyone else's results look absolutely amazing). Really not sure what you're doing different with their models, but whatever it is, it somehow makes it work much better for my prompting style and settings. Ease-of-use, which makes things more fun imo!
Other than having to externally load a VAE for Test_ v5.0 (which I often forget to and have to start some batches over) and the fingers/hands issue, it works very well and produces some amazing results. Might be just a slight bit too stylized compared to v4.0, but they're both very nice, just different. Kinda feel like a simple 50/50 merge of both would be a little more balanced for my tastes (or maybe ~60/40 heavier with v5.0), but I've probably grown a little too partial to v4.0 over the past few weeks.
The software I'm using finally got an update to support Flux, Qwen, ZiT and newer models, so I'll probably be experimenting with that stuff for a little while you're working on your anime branch (I really like watching anime and cartoons, but my personal interest in using AI has always been with photorealistic stuff). Unfortunately, afaik, this new update doesn't change VAE loading behavior and I still gotta remember to manually select one for each prompt if it isn't baked-in to the model I'm using. It's not a deal-breaker or anything, just a pain in the butt cause I forget to all the time, lol.
Anyway, very excellent work with v5.0 and thank you so much for sharing this with us!
And to add, wow. After playing around with Test_ v5.0 a little more, it is quite nice! Definitely feeling the added creativity of wildzzz stuff, with its elevated artistry and expression. Think it trades off just a touch of IQ/realism in the process, but the depth, illumination, composition and prompt/LoRA adherence is nearly flawless imo (near-perfect character likeness, can throw a dozen LoRAs and max tokens, and with enough steps it still manages to figure out everything just fine). It's a little wild and raw, but quite fun and capable!