santa hat
deerdeer nosedeer glow
Sign In


Checkpoint Merge
Aug 15, 2023
Base Model
SD 1.5
iamxenos's Avatar
Joined Nov 28, 2023


It has come to the point where I've managed to get XenoGASM so that it's putting out as real - or maybe even more realistic - generations than this model. Plus - that one is a lot better at other concepts than this. As such - XenoREALITY will likely cease to be updated down the road.

I'll leave this here for folks to use if they want - but if you're looking for something that will continue to be updated and improved every couple of months - head over to the XenoGASM model and give it a shot.

That's Not Real, That's Xeno Real!

This model represents my first foray into more realistic models. It is designed for both portraits and scenery. At this stage, it is not particularly great at scenes with people in extreme positions - but I think I'm on the right track. It's a lot better than my last few methods of achieving this goal.

This model is not inherently capable of porn (though all of your LoRa's should work - please do share your successes and failures so I can get an idea of what needs the most work.)

This model does tend to go nude fairly easily, though.

If you test and play with this model, please show me some of your failures and successes (and please don't just tank the review ratings without at least some clues and information I can use to try to fix the problem). There WILL be issues compared to many of the more mature and stable models, but with your help I can make this great and provide some unique things not seen in other photo real models.

Like it? Buy me a beer. * Shoutout and Thank you to Arpigeon for the beer! I will be enjoying it and sending you a toast during my Friday Happy Hour this week!

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.

V2.0 - True Grit

V2 should push even harder toward "realistic photography" while still taking on characteristics of various art styles and so on. While it may be difficult in some cases to get an art style looking like drawn art - it WILL tend to take on other elements of that style (shaping, outfit selection, etc) while maintaining a "real" look and feel to things. Your default people and their looks will be more raw and gritty, while still hopefully being attractive (and you can always "hollywoodize" everyone with various beauty and attractiveness statements too).

RC 1

From the very beginning of this learning process (with my XenoEngine models) I've been working on trying to create a versatile model that will work well on "Potato" computers (i.e. LowVRAM and NoGPU environments) in order to create game character assets for games. This means we want to be able to generate consistent characters in different outfits and styles without big long prompts and without needing to make 20 attempts just to get something decent or that doesn't need 3 hours of post work just to get ready to start doing background removal and optimizing for game use. This would be for 2D game engines like RPG Maker (character busts), Twine, RenPy (and other visual novel systems) and so on.

This Release Candidate represents what I think just might be very close to having accomplished that. These last optimizations work really well (most of the preview images I'm going to post to start were done on a No GPU Huggingface Space). Memory optimizations seem really good - and I'll be posting some tips and recommendations over the coming days as we run what I hope is the final round before an official release of the model.

Some key changes in this:

  • Memory Optimization: All along, I've been trying to test (and lately, I've even been performing the merges themselves) on both T4 GPU environments AND No GPU and LowVRAM systems. XenoEngine ArtStyle never really worked that well for the Potato Computer systems, but I still love it for higher end installs. Each of the next two (XenoEngine 5th Dimension and 6th Sense) got a little better at it. This one... well, it looks pretty good (until you try to jam too much action and specific elements to include in an image). And, of course, it looks even better for these situations on a proper rig.

  • More Outfit Diversity: While I feel that the outfits that come out when you just generally describe them (e.g. "Business Suit" or "Bimbocore Outfit") are less "sexy" by default - the potential variation while still saying on theme is greater. (And you can still get to those sexier outfits by being a bit more comprehensive in your description of what you want them wearing.)

GPU vs. NoGPU Examples:

GPU Examples

NoGPU Examples: Random Female Celebrities / Random Male Celebrities (Soon) / Various Random Art & Composition Styles


This is an update that is mostly sanity fixes - better default camera positioning, better fallback lighting and composition when you don't declare anything, and a bunch of new outfits. As you can see from the comparison images, the changes are subtle for most complicated prompts with lots of direction - but simple prompts have a much greater change.

Concept Blending is working really well and over a much wider range of concepts now, too. Here are some examples with a brief description. it's REALLY cool and fun to play with if you haven't done it before.

Main Goals for Model

  • Improved Character Sculpting Controls: One of the things I love about the comic/cartoon based models is that they tend to have much more effective and comprehensive body controls - but the reality based ones are much more limited. It makes sense. For example - source images that show lots of different breast sizes - especially ones in that show comparison, differences, progressions, and high range sizes are digital art coming out of the fetish community. As such, as soon as you start applying those changes, your image starts to shift away from real toward the style of the source images. The bigger the shape change, the more the style is going to change.

    I've developed a recipe, though, which allows me to get in (some of) those controls with only moderate effect on things. At this point you can make some small to moderate changes without much change. Larger changes start to move it toward looking like a 3D render a bit more than a photo. The main issue now is a loss of skin detail - which may be able to be brought back with some clever prompting (let me know if you find something) but I'm also working on it from this end, too.

  • Smart & Consistent Default Fallback Tendencies: When we run a prompt that is missing something like a lighting, scene description, camera angle, and so on - SD will use it's (sic) Intelligence to make that choice for us. In my other models, they did a pretty good job of this - but all the tendencies and things I put in there were all based upon things that relate to the overall drawing/rendering style - not lighting and quality/reality tags and all that stuff.

    Right now, if you just put a subject name and a simple action or description, the output is just sort of flat and boring. An early experiment I did got some nice detail and effect going - but it wasn't consistent. It might choose a nice golden hour in one shot and then, same prompt, different seed, light it completely differently. That's not what I want - I want it at least somewhat consistent for story tellers and game makers.

    For now, make sure you're including some lighting, style/quality, and other appropriate tokens to get that "finished look" going. Those triggers won't change, but hopefully I can get it so it's making good choices so you don't NEED to put all the gazillion tokens in there just to get a halfway decent picture.

  • Lots More: There are a zillion things this model already does well - things that you don't necessarily see very often in realistic models. I want to get to uploading this thing so you can play with it though.

    Look for updates on this page (even when the model doesn't update) and I'll also be posting some articles with tips and tricks and all that fun stuff.

Again - this is BETA, but still already slightly above average for a checkpoint, IMO. Don't expect perfection, yet, and please share issues, and problems (and successes, too!) so that I can address them - either by helping you figure how to prompt your way out of it, or to make adjustments in the model itself (if possible).