From V3 : Switched to 4x-Ultrasharp as upscaler
SR mean Semi-Realistic
I already made a2.5D and anime model, so i wanted a littlemore realistic one (but nof fully realistic... Not my thing XD).
So here the result. Not bad i think ^^
About rhis Model
This model is capable of, of course, generating cute and sexy girl, from petite and skinny cute one to tall and muscular beautifull one. You like enormous breasts? it can do it. You like big ass? it can. You like some meat around the bone? let's eat! you are more on the the "flat chest is justice" well, be my guest!
Men are also an option.
Background wise it's fine too.
Ofc, is Ecchi/NSFW/Adult content is avaible too. But for more specifics kink LoRa are advised.
Oh and ir's LoRa compatible : anime and realistic one (but it's better with Realistic LoRa).
The best way to use this model are to do portrait image. You can still you things from further but, from cowboy shot, the face details and hands start to become a little messy.
The merge details are in the About this Version. The VAE is already baked.
All the Sample were generated with the fp32 Version.
Generation Settings and Prompt.
Sampler and Sampling steps : DPM++ 2M Karras at 40 steps or DPM++ SDE whit steps betwenn 20 to 30.
CFG Scales : from 6 to 8. Depend if you want less realistic or more realistic.
Upscaler : ESRGANx4
Positive Prompt : i mostly use (masterpiece, best quality) is enought for getting good things. You can other tags like realistic, photorealistic, hdr, cinematic lighting, anime coloring, pastel colors etc.... To get different output. Just avoiding tags about the resolution like hughres, absurdres, 4k, 8k etc... white generating images further than face or upper body. The eyes will become pixelated.
sexually sugestive and nsfw tags lewdify a sfw prompt from sexy to total nsfw. It's a lazy way to do something lewd with a simple prompt.
Negative Prompt : i used the following (worst quality, low quality), lowres, watermark, signature, logo, monochrome, greyscale, jpeg artifacts, aliasing, blurry, resized, duplicate, error, what, bad anatomy, bad proportions, bad perspective, bad hands, deformed hands, bad fingers, deformed fingers, fused fingers, interlocked fingers, bad feet, wrong hand, wrong feet, extra digits, fewer digits, extra arms, extra legs, missing limb, disembodied limb, thick eyebrows.
Mostly for fixing hands without the use of TI/Negative embeddings, and to prevent som common errors to appear.
This a basic negative prompt so, use what you want and find you own way to generate :).
If you want to use TI/Negative embeddings go ahead. Just avoid to use the EasyNegative and Badhand V4.... They just screw the model's style.... And not a little.
Conclusion
I hope you'll like it and have fun with it.
Don't hesitate to post here the images you generated with it (even the hard stuff) it will make me quite happy to see people using it, and, i will probably discover some really nice way to use it.
Enjoy! ^^
Description
Merged Model
Baked Vae : vae-ft-mse-840000-ema-pruned
FAQ
Comments (7)
Hey, well done on a great mix!!
I really, like her, I think she's gonna grow on me!!
@kellykelly
Thank you :)
Don't hesitate to post you generation ^^
I like semi-realistics.
At first, just looking at your demo gallery, I must say the so called "SR" feature should be a bit improved. The first image is great but all the other are more "anime" than Semi-Realistic...
I will try the checkpoint to tell more...
@settima_ai It depend of the hair color.
If you use black, brown or blonde hair it will go more on the realistic side.
The others hair colors lean a bit more on the Animé side
If you want to use pink hair, add realistic tag if it's lean a little too much on the Anime side.
Thank you for the feedback :)
@Lorim I've tested and there are some good points: solid graphic and very precise on arms/legs. Success rate is high, it will go wrong only with complicate prompts. I've tested also various LORA with no problem.
One point I dont like much is the tendency to drive faces to a certain "japanese" standard (I call it this way because it remembers me some checkpoints from a famous japanese user)
Even if I use a TI or LORA of a character, I get the right face but, looking with more attention, there will be always a tendency to drive eyes and face to that "japanese" standard. (just a bit, not really a big problem)
@settima_ai I'm aware of the model tendencies to lean a bit too much on the asian side.
i'm trying to fix that and to get a more "neutral" face.... But it's a bit hard XD
Because i know where it's come from, and i daon't want to remove this model (NeverEndingDream) from the merge ^^;
@Lorim A 3rd image has been release from "approvation". Dont know why this one was blocked for inspection, since is very similiar to provious two I posted...
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.





