All images out of the box, no postprocessing. Model
works best with high sampling levels. Larger resolutions
are also working good. If you have enough vram sometimes
no upscaler needed.
Recommendations (default use):
Optimum CFG-Scale: 7
Optimum Sampling Level: 150
Sampler Best Results:
Euler a
DPM++ 2S a Karras
DPM++ 3M SDE
DPM2 a Karras
Restart
The model is highly optimized on the settings
above, so you should leave out quality prompts
first, if not necessary.
The model is build of the latest trained models
and lora, thanks to all community members
for their work.
I hope you enjoy and share your images.
Ultrium is always a merge of the best and/or latest
avaiable models. So each Ultrium is different than
its predecessors. It's your decision if or which
Ultrium is best for you.
Ultrium 9.333 will be the last Ultrium for SDXL 1.0,
I will see for which AI I continue my work.
Description
new Version
FAQ
Comments (28)
Hey, I think you uploaded the fp32 model twice- they're both 12GB.
Thanks for the hint, I deleted one and the pruned version should be available in a while. (Looong upload).
a merge ?
What are the differences and advantages of the other models?
It is a pity that you do not mention the owner!
Hello,
generally a trained model can be better (my opinion) if it can do all the prompts you have in mind. If not, you use a lora f.e. so gain this specific ability for your prompt. With a merge (model + model or model + lora) you can implement the abilities directly into the main model. But there are sideffects to f.e. the quality, sharpness, text. Ultrium 5 has around 30 loras integrated and consists of many models, so there is a lot this model can do out of the box, but it becomes more and more of a task to balance and ensure the quality.
My goal is to balance the integrated content so good, that there is no need to invest a lot of time to tune the lora-weights for each image. If the tuning is imperfect you lose sharpness, have weird hands or messed up text f.e.
There would be no use list the content of the model, because nearly every trained model and lora have been integrated.
oh thy for explanation ..
may you intersting in my loras ... blinking eyes if it works to you ;)
Couldn't run it on Fooocus
Sadly I don't know focus, so I can't give you any hint. Have you tested full and pruned?
V5 is pixelated, overexposed, cartoony, and has problems with eyes in comparison to V4. V4 is much more natural and realistic looking. Sorry to be so critical - on a positive note, V4 is the best NSFW model I've used!
I posted some pics to demonstrate.
Hello,
thanks for your comment. As always it's still a process, some thinks improve (f.e. the level of detail of 5 is sometimes much higher than 4) and other things get worse (f.e. too much pussies, more stubborn for vertain prompts than 4 and things like anime not working consequently). I'm aiming to regain some qualities of U4 and keep the detail of 5.
I like this model a lot but my renders lack A LOT of detail. The model it's self is awesome. I just wish my characters weren't fuzzy. It's very granulated. I've tried skipping the first 2 steps. Different samplers. Different prompts including the ones you suggested. Putting emphasis on low quality in negative prompts. Nothing seems to work to get the amazingly good quality. I get perfect quality from Dynavision for example. Any suggestions would be nice. I use Comfi-ui. Thanks
Hello, have you tested "8k, epic realistic, ultra detail"? Negative prompts like "deformed" or "bad hands". CFG-Scale should be around 0.7-0.9 and I suggest Euler A as renderer. It did not test Comfi-ui - so it could be related to this as well. Your are right dynavision is great, it is one of the cleanest models in the aspect of output quality - but it has a little limited content :)
Many people are waiting for Tile model for SDXL, if I'm not mistaken at the moment it doesn't exist. When (if) released it'll give you as much detail as you need. Another trick you could use is applying another model as a "second run", it can be during upscaling (including with x1 factor, meaning without changing the image size). Of course, in this case you cannot set high denoise value and it may still lack details you need and you probably need a model that kind of similar to this one. In any case it requires a lot of experimentation (well, like the whole SD thing to be honest).
What about sharing your works and prompts? Maybe we can help you, get better images. Just saying, it isn't good, doesn't help you or the creators. But i am sure, a lot of users wanna help you :)
You could also try "Detail Tweaker XL" LoRA - https://civitai.com/models/122359
Hey, so I've been playing around with this because I know what you mean. Faces are pixelated.
Try rendering in 768x1344. When viewed on a 1080p screen, they will no longer be "pixeleated". Consider it downscaling. DPM++ SDE (not 2M or 3M SDE) is the best sampler for this model. Euler A lacks detail and 2M/3M has this weird mosaic texture.
Negative prompts and "8k, ultra detail" had no effect. The only prompt that improved quality was "cinematic lighting". Try moving the subject to the end of the prompt as well, as I think the model is overfitting when you have the subject at the beginning of the prompt.
It seems that the v5 model has some strange overfitting on "shorts", it tends to generate some bikini-like underwear and expose much of groin rather than normal sfw ones
Hello,
this is related to the bleeding of prompts, if you train bikini by lora f.e. this will (sadly) also affect other clothing. This is also the reason why U5 can't do male nudes - training this would lead to woman with hairy legs or transgender. The difficulty is to balance almost everything in a way, that nothing is over weighted. And even if you take care there can be effects you don't realize until someone tells you (f.e. the model looses the ability to do night-szenes or it "forgets" some words.
@DSchroeder If you train the model yourself, you may introduce some model generated image as a balance (also known as "regularization set"), for example, if you train some bikini image, you can use captions with "bikini" removed as prompts, generate some normal cloths with the base model, and add them to your training set. This will give trainer a "force" to stablize other concepts other than "bikini".
This was confusing. it came up when I searched for VAE. I was looking for a standalone.
Maybe you can rename next release as "... + bakedVAE" or something?
Thanks for the hint, I will consider this :)
I don't use Auto1111, can someone please convert this to Diffusers? I'm getting an error on all versions no matter which one I download, and V3 seems to download V5
(Edit: this was written before Ultrium v6, which I haven't tested yet, 11/30.)
I just ran an exhaustive comparison of 55 NSFW photorealistic SDXL checkpoints looking for things like prompt obedience, anatomical correctness, and various aesthetic factors, and found that House of Freyja - Ire's Mix was the pretty clear #1.
Runners-up included Better Than Words v20 and v30, Capability XL, Luna Mia v50, and Ultrium v4, in no particular order.
yeah, we can see by the examples images the creator himself posted how "anatomically correct" House of Freyja is...
what about onlynsfw xl?
@jackluz in all seriousness i've learned not to judge by author's examples because even they may not prompt as well as others could with their own checkpoint. Freyja is much better for me than the author's examples. I agree they are bad example images. Another checkpoint that exhibits this is Hassan's SDXL NSFW. His example images are terrible and look nothing like what myself or others can achieve for outputs. https://civitai.com/models/207597
In reality, ranking checkpoints will always be highly subjective. I tried to be as generalized as I could but it's just not going to match everyone else's experience. As I've said elsewhere, you just have to find a checkpoint that "speaks your language" (kind of literally). Language itself is very subjective.
@Picky2011670 I did test onlyfornsfw and found it quite lacking in the "pretty" department especially when explicit content was part of the image (could be ok as long as a cock or anus wasn't in the image, otherwise the face just went to crap), and it also didn't obey my prompts very well. It was one of those "outlier" checkpoints that tended to produce something entirely different when most other checkpoints were producing obviously-similar stuff on the same seed. It ultimately ranked pretty low on my list, to be honest. I still have it in my carousel of checkpoints on my drive but i almost never use it. I've kept it around just as a future option if I get into merging myself
which one can produce the best male 'anatomy' that seems to be the hardest thing to make checkpoints do correctly.
@redbible largely the best option is to use one of the many penis LoRAs available for SDXL but if you really want to get it from the checkpoint itself, you could go with Himeros XL or LimitlessVision XL. But once the supply of penis LoRAs became adequate about 1.5 months ago, I stopped keeping track of which checkpoints worked for them, so if there's a newer checkpoint with solid ... uh, penis performance... I might have missed it.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.