What began as an effort to breathe some new life into some older models has evolved into this little flagship quality checkpoint. (More on the backstory below... for those who give a damn.)
XenoGRAPH
This is, by far, the most photorealistic model I've created to date. While I was originally thinking of this being the next version of XenoREALITY, it ended up being so much more than that - so it's now it's own new fork.
Realistic Photography: This was the main objective for this checkpoint. I'm not a prompt master, so I'm sure some of you token aficionados will be able to do better - but I've landed on the following as a good lead-in template to get some pretty realistic looking photos.
Masterpiece, Best Quality, ultrarealistic photography, real life, colorful, amazing, perfect lighting, bright colors, dramatic, dynamic, cinematic lighting, hyperrealism, intricate details, epic, realism, That said... it typically does a pretty good job of being at least fairly realistic without any help at all.
Realistic Fantasy/Sci-Fi: This was the big surprise. XenoREALITY was always really good at bringing other art styles into it - but this baby really shines when you mix in artistic elements and fantasy/sci-fi themes. I'm not even sure how to describe it - it's sort of like fine photography then put into the hands of the greatest photoshop digital artists and finished that way. Or something. Check some of the fantasy/sci-fi themed preview images and you can decide.
Who Knows What Else? I've been running test batches and wildcard combos to put this through its paces for two days now - trying to figure out what this model is best at. Every time I think I've got it figured out... a new token set or concept sends me down a different road and it becomes a whole new model to explore. (I've been especially keen on "<fantasy creature/thing> transformation" as a fun token. It does a really nice job of catching a midpoint during the transformation with a very high hit/miss ratio.
Generation Tips
I did most of the testing with Euler A or LCM Sampler (with LCM LoRa enabled). As with changing between any samplers, there's a difference, but generally the quality and feel come out great with both. Your own favorite sampler should do a fine job as well.
Complicated prompts? Sure... it seems to love them - and I've found little tiny hints of things which, on first glance, might seem like it missed from your prompt, but then you realize that there's just a little hint of it somewhere.
Simple prompts - yep. Just a name and maybe some clothes or whatever generates a really nice picture.
Negative Prompts: This is the first time I've gotten through a testing of a checkpoint where I didn't feel like I needed to toss in some negative prompts to help it along. Bad hands - sure, once in a while, but not so often that I want to put a badhands negative embed in as a default. If you check the preview images, you'll find that almost none of them have ANY negative prompt whatsoever.
Portraits: Of course. I can't imagine releasing a checkpoint that didn't do great portraits.
Scenery/Backgrounds: Yep. And some really cool stuff, too. Mixing concepts creates some really fun environments.
LoRas and Embeds: Oh yeah! And I've found that it especially likes a bunch of LoRas and Embeds if you go into LCM Mode (using the LCM Lora and Sampler - though I'll release a pre-made LCM version later this week for those who want it). In one early stress test, I had 6 different LoRas going (counting the LCM) and it came out excellent.
My best advice for trying out this checkpoint - just do you and see what happens. Use the sampler and settings you like - and go for it! Can't wait to see what you come up with!
The Boring Backstory
Once a checkpoint has been merged and remerged enough times, you start to hit a point of diminishing returns when trying to update and add something new to it. Many concepts get over-trained and other concepts just sort of fall out of the bottom forgotten as you pour more new things into the top. Both my XenoREALITY and XenoGASM checkpoints have reached that point... anything new makes it lose something else.
But I was really wanting that one great "this is super realistic" model, so I started playing with various "trained" checkpoints that focused on photography and realism. And honestly, the mixing and merging experiments went on for about 4 days and I'm not sure what all is in there. I know "Subreddit V7" was a key factor because I like the way it takes some of the polish and glam out of things - your default 1girl is more "that girl from the office" and not "some supermodel".
Eventually, I ended up with something that was a great mix-in - or at least that I thought would be. So I merged it into XenoREALITY and.... was fairly unimpressed. It was different. It was probably a bit more realistic... but it just didn't have any magic.
So then I scrapped that idea and thought maybe it would work well as a fork of XenoGASM... so I went back to some older versions and tried merging my mixer model in there. Same effect. The one merged into V3 was the best... but it still lacked that special something...
Finally, on a silly whim, I wondered what would happen if I then merged my lackluster XenoGASM and XenoREALITY failures together... so I tried it and... WHAM. This is what we got. (Actually... that's not exactly true - once I realized it was freaking awesome... I merged in a little 0.2 weight of Xeno Detailer to help clean up the eyes and hands a bit).
And so... this isn't the next XenoREALITY nor is it the next XenoGASM. It's a science of its own - XenoGRAPHY, if you will.
Have fun and I can't wait to see what all you come up with!
Want Tips, Tricks, Pix of Things We Can't Show on Civitai?
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.
