Find XenoGASM-MK2 Here!
Future updates to this model will be ceasing. I can't get it any more real without losing more concepts and I can't add more concepts without losing realism. So... I've reworked things and come up with a new fork that makes MASSIVE strides in making XenoGASM the coolest checkpoint ever.
Checkpoint: https://civarchive.com/models/315459
Information, History, Tips: https://civarchive.com/articles/4206
Based on the XenoREALITY model, this one has most of the more NSFW features reconnected and fine tuned to encourage more "sexy" outputs. I wouldn't go so far as calling this a "Porn" model - but it could definitely be a good base for your forays into that sort of subject matter.
Join the XenoEngine Discord to discuss tips and tricks, complimentary Lora's and Embeds, and to post (and see) images not allowed here on Civitai.
If you like this model, please take a moment to leave a review.
LCM Version Available! HERE - Learn more about LCM Here
Article: Using SD 1.5 Checkpoints as Refiners for Enhanced Realism
V5 - Primarily retuning and optimizations. (See version info to the right)
V4 - See Version info to the right
V3
This version managed to push the default outputs MUCH more toward photographic realism as compared to the "semi-real" outputs from earlier versions. Still - the more kinky/fetishy you get, the less "real" that default output will be and the harder it will be to actually prompt it back toward being "super real". Even so - that push seems a lot less intense in this version than previous ones.
There are also some cool things going on in terms of racial and general diversity among characters. <adjective> woman (or girl, or whatever) will change the default "woman" into someone else - often to another race. Try things like "sexy girl" vs. "beautiful girl". (examples)
RE: v2 - Dark Evil vs. v2.1
This v2.1 is the version that was supposed to be released as v2.0. Since I'm an idiot, I uploaded the wrong file for that. THIS is the one you want... probably.
The v2.0 I posted yesterday is still up - it doesn't belong being the next version of this model, but there are some fun, if not dark and disturbing, things happening so I'll leave it up for those who want to play.
General Tips and Information
Start w/ Simple Prompts: This model opens up a LOT of token connections which normally don't exist. In other words, when you say <this word> it will connect to extra details and thematic elements that are related to the word you used - without needing to write it all out. The upside is that you can get very robust images with very simple prompts. The learning trick is that if you get too much in there, you can get too much going on and you'll need a lot more iterations or just generally simplifying your prompt in order to make it all work. (Once you figure out what works for what and how controlling this model are different - your template prompts can more easily come into play for use once you trim out and tweak them accordingly.)
Keep in mind, also... "short prompts" is mostly about the action and what you are actually seeing. Your camera, quality, and style info - though it may work somewhat differently than on other models - that can typically be longer as needed.Monster/Creature v. Damsel Type Renders Can Be Tricky: Because this model allows (and encourages) us to be able to make some pretty creative monsters and creatures out of our subjects, a lot of the pathways between "This is a monster" and "This is the Female Subject" are opened up. As such, when you describe 1girl and a monster - it will tend to make our girl BE the monster, not be the victim of said creature. It can be accomplished with inpaint and potentially with ControlNet and/or After Detailer (still running some experiments on the latter two) but you may have trouble with a single-prompt, single pass, no-postwork type scenario.
Lower CFG and Denoising For Fantasy/Non-Realistic Things: Generally, the default settings and your choice of Sampler and Upscaler will make a "difference" but they generally all work as you'd expect them to. That said, this model, at it's heart, wants realism. And then some jerk (namely me), taught it to do all sorts of unrealistic and downright perverted things. With your denoising/CFG Scales and all that set too high - then you'll often find that things like glowing eyes, pointy ears, emissions of sexy mind controlling pheromones, and things like that will end up getting washed away (especially during the upscaling process). Lowering the CFG/Denoising a bit (5-6 for either - though lowering the Denoising on Upscaling seems to have the biggest impact of the two) help keep the model from saying, "Hey, that's a woman - women don't have pointy ears, glowing eyes, and weird mind controlling particles emitting from their bodies."
Go Easy On Negative Prompting - Especially for Fantastic Style Things: Again, this model is a sort of mix of things - it's a (mostly) realistic look while generating things we often only see in Anime/Hentai and other Art Based Models. So if you are using a negative prompt designed for realism, a lot of the "fantasy" type things may have trouble showing up. Accordingly, if you use a negative prompt designed for art/Anime and kinky stuff - you may lose a lot of the realism you're trying to achieve. I like to start empty and add things into the negative prompt as needed rather than starting with a bunch of things that I assume I'm going to need.
Check the notes for Xeno Engine Artstyle and XenoREALITY for more tips. At the base, the core features are the same (though those two don't have the new MC additions yet - ArtStyle will be getting those things next week). Each model outputs a bit differently, but you can tend to punch and prod them in similar ways to get similar results. If you visit our Discord, You'll also find lots of great tips, see images that would be frowned upon here, and get lots of help and advice in getting the most out of this and the rest of my models.
Description
This is the version that was supposed to be released as v2.0. Since I'm an idiot, I uploaded the wrong file for that. THIS is the one you want... probably.
The v2.0 I posted yesterday is still up - it doesn't belong being the next version of this model, but there are some fun, if not dark and disturbing, things happening so I'll leave it up for those who want to play.
For general usage and to follow the proper upgrade trajectory... this and future versions will be the way to go.
FAQ
Comments (14)
You say you can use low numbers of words and Go Easy On Negative Prompting get good results.
But your posted pics say (see below) and look at all your post mods you do.
photo RAW, portrait,
parasite infection, clear slime, tentacles, multiple girls, female orgasm, female female ejaculation, wet skin, covered in clear dripping cum,
masterpiece, award winning photography, lighting, perfect composition, high detail, hyper realistic,dramatic lighting, epic
Negative prompt: SkinPerfection_NegV1, (worst quality, low quality, normal quality, lowres, low details, oversaturated, undersaturated, overexposed, underexposed, grayscale, bw, bad photo, bad photography, bad art)++++, (watermark, signature, text font, username, error, logo, words, letters, digits, autograph, trademark, name)+, (blur, blurry, grainy), morbid, ugly, asymmetrical, mutated malformed, mutilated, poorly lit, bad shadow, draft, cropped, out of frame, cut off, censored, jpeg artifacts, out of focus, glitch, duplicate, (bad hands, bad anatomy, bad body, bad face, bad teeth, bad arms, bad legs, deformities)++
Steps: 20, Size: 480x640, Seed: 1173304724, Model: XenoGASM-v2, Version: v1.5.0, Sampler: Euler a, CFG scale: 7, Pad conds: True, Hires steps: 10, Hires upscale: 2, Hires upscaler: Latent, CFG Rescale phi: 0, Denoising strength: 0.75
Those examples are pushing hard at the upper limits of prompt complexity - trying to show as many different things going on in each image as possible (without everything going to crap). The extra style calls in each batch also help to generate a bit of consistency in look and feel across the other images in the set. <shrug>
They aren't so much about showing great prompts to make great art. They're about showing prompts that generate things that this model does that others might not (or that they might do differently). Also - once you find a style prompt that "works" - it will tend to keep working. And from there you can create your template.
My warning is more about not STARTING with a big prompt because you can't always be sure exactly what part of it is triggering what effect (an effect you might desire, or an effect you want to get rid of, either way). Build out a good prompt and from there you can start building templates for specific things to keep tone consistency and whatnot across generations within a set.
At the end of the day though - prompt how you want and how you do to make great art. But there ARE a bunch of things in this model that work differently from most, so starting with simple and building is the best way to learn and to teach you to better cause or prevent those things from happening next time.
With the "Celebrities Walking Down The Street" prompts too - if you look, the "style" part of it is the same, but the actual "prompt" that describes who is in the picture and what they are doing is very simple:
carla gallo atompunk daywear outfit <--- A name, a style, and the fact that we want a "daywear" (or "eveningwear" in some) outfits.
@iamxenos I am sorry I missed the helpful part suggesting not "starting" with a lot of prompts, that is very helpful, thanks. :)
@pursuit_of_beauty More importantly, I'm sorry for making it confusing. I reworded it a bit, but it still probably needs some work. Either way... I appreciate the feedback. Cheers and happy generating!
You see this mindless prompt vomit everywhere. Sheeples don't even realise that you get token 6323 for both "raw" and "RAW", and yet they still enter it in capitals lol. You can kind of understand it from the average incurious person, but disappointing from model creators. With decent models it's really best to start with nothing, maybe use just "person" to get some idea of flavour, add a few words that you know from testing are actually effective and that you always want, although even stalwarts such as "asian" in the negative are no longer always required for non-asian output, and go from there. "If it aint broke don't [try to] fix it" is very much an applicable maxim.
@nickfli121 The showcase image has more stuff, but the rest are pretty simple.
And yes, I use RAW because that is correct, not because raw doesn't work. I capitalize people's names (almost) all the time, too. <shrug>
@iamxenos We're all creatures of habit :) Though FWIW, neither "raw" nor "RAW" work. You might try comparing the result with and without that word, or with a multiplier of 1 and then 0, to see if it's doing what you want. Also given that raw is an adjective (which models like), and in some contexts would be associated with something visual (e.g. raw flesh, a gritty scene), if it influences at all it might be more likely to be as an adjective. RAW images are also almost never found on the web outside of people sharing them for demo purposes, camera reviews, discussions of problems with processing them and the like, so I'd imagine the amount of RAW images in the training data (if any whatsoever) would be low (would be very interesting if >10% of their data was but I seriously doubt it). I do find words such as "perfect" can be useful introducing a prompt. I came up with adding "sharon osbourne" to my negatives and it's a current favourite. Film stocks tend to work just on their own with no other words necessary. "arri" gives a slightly different composition to "xt3". Lens specs such as "85mm" almost never do anything (it might appear that do at first but in testing you find they actually don't). "long lens", "short lens" tend similarly to be ineffective, "fisheye" does work and "wide angle" can. I just think some care when giving examples is a good idea.
@nickfli121 There are some good points in there, but I'm going to have to disagree with RAW not doing anything.
AKAIK, the "RAW" that SD was trained on isn't the image format, but it's not really an adjective either. It's for images that are in raw form and haven't been filtered, color balanced, compressed, or otherwise post processed - thus giving them a more raw look. (Whether I choose to capitalize it or not).
In the following pair - RAW definitely makes her less "finished" looking - more human, and less "fixed for glamor". The jeans are rattier, the skin has a few subtle defects, and so on.
https://civitai.com/images/2680250
It may not do a lot, but to say that I'm creating "mindless prompt vomit" is not an accurate assessment at all. Sure - sometimes I leave in a token or two that could have or should have gone away from gen to gen - but I don't mindlessly slap words into a prompt without knowing that they do SOMETHING - and that that something is something I want to see.
@iamxenos 1.5 was trained on LAION-5B, and I couldn't find any discussion of "raw" used in the way you said in the dataset, but I'd be interested in any references about that. It could be a myth though. I certainly agree on the difference in your image test, which is what I was alluding to with raw as an adjective, with the roughed up jeans giving a more gritty less refined image in terms of content. I didn't pixel peep or throw the image into a histogram or any scopes, but the quality in other respects seems identical though, with the same dynamic range, tendency for and proportion of hotspots and so on. Doing a search of the training dataset online for "raw,jeans,person" shows images with people having ripped jeans as you'd imagine, hence your result I believe.
https://rom1504.github.io/clip-retrieval/?back=https%3A%2F%2Fknn.laion.ai&index=laion5B-H-14&useMclip=false&query=raw%2Cjeans%2Cperson
@nickfli121 The quality in terms of what you're talking about will be the same because both images are JPEG files - there isn't a RAW file in there to make a difference. No tokens do things like that - when you type 8K it doesn't make the image 8K - it just makes it refer to vector references tagged that way which might also coincide with other tokens in there.
@nickfli121 And really - that's not the point. The point is that you came in here calling me a prompt vomiter because I used RAW in a prompt and - according to your wisdom - it does nothing. Which it does. It doesn't matter what it does - what matters is that you are rude and really should apologize for being so.
Is v2 less 'semi real' than v1? Looks more artistic than v1, but maybe just more flexible?
I'll have to read through what I wrote and clarify that description when I find it, I guess. What I meant is that it's more "semi" than "real" - in other words, we've lost a bit of realism in favor of being - yes - more artistic and flexible. I write faster than my brain works sometimes. lol Sorry about the confusion there.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.