What it does
Xenophage is an SDXL concept LoRA trained on Illustrious for alien/symbiote corruption, infestation, and transformation themes. Use it for body horror, sci-fi horror, biomechanical environments, egg-infested scenes, and full-body alien transformation imagery — from minor neck-and-arm corruption to full hive-bound transcendence.
Built around a curated 999-image dataset with rigorously hand-tagged stages, body-location specificity, and 11 color variants.
Triggers
Core: xenophage
Parametric — substitute [color] and [bodypart]:
[color] alien corruption,[color] alien eggs,[color] glowing eyes,[color] glowing tongue,[color] streaks in hair,glowing [color] veinsalien corruption on [bodypart]Stages:
minor / partial / full alien corruption
Colors: blue · purple · green · gold · teal · red · silver · black work strongly. Pink and white are weaker (smaller training share). Rainbow tends to read as purple — boost with (iridescent, oil slick, holographic, prismatic:1.2) to push toward true multi-hue.
Bodyparts: face · neck · chest · torso · shoulders · arms · hands · legs · feet · back · tongue
Environment vocabulary: alien tendrils, alien tendrils emerging from egg/slime, alien egg cluster, alien hive, biomechanical, infestation, dripping slime
Good at
Stage-controllable transformation (the minor → full progression is a real axis, not just intensity)
Anatomical specificity (corruption stays where you put it)
Egg/tendril environments without a character (use
no humans, sceneryfor still lifes)Stacks cleanly with style LoRAs — tested across multiple style mixes without degradation
Multi-character and pose flexibility — concept doesn't override scene composition
Recommended settings
Weight:
0.8 – 1.0Works at any SDXL resolution; trained at ~1MP bucket sizes
Clip skip: 2 (Illustrious standard)
Best epoch: 25 (clean, no overtrain)
Xenophage is very flexible with style — tested cleanly across multiple style LoRAs without quality degradation or concept bleed. It also stacks well with other adjacent concept LoRAs to push their output toward the xeno aesthetic: try mixing it with venom symbiote, oviposition, tentacle, body horror, or biomechanical LoRAs to give them a corruption / infestation flavor. Recommended weight when stacking with another concept LoRA: 0.3 – 0.8 to let the partner LoRA lead.
Description
What it does
Xenophage is an SDXL concept LoRA trained on Illustrious for alien/symbiote corruption, infestation, and transformation themes. Use it for body horror, sci-fi horror, biomechanical environments, egg-infested scenes, and full-body alien transformation imagery — from minor neck-and-arm corruption to full hive-bound transcendence.
Built around a curated 999-image dataset with rigorously hand-tagged stages, body-location specificity, and 11 color variants.
Triggers
Core: xenophage
Parametric — substitute [color] and [bodypart]:
[color] alien corruption,[color] alien eggs,[color] glowing eyes,[color] glowing tongue,[color] streaks in hair,glowing [color] veinsalien corruption on [bodypart]Stages:
minor / partial / full alien corruption
Colors: blue · purple · green · gold · teal · red · silver · black work strongly. Pink and white are weaker (smaller training share). Rainbow tends to read as purple — boost with (iridescent, oil slick, holographic, prismatic:1.2) to push toward true multi-hue.
Bodyparts: face · neck · chest · torso · shoulders · arms · hands · legs · feet · back · tongue
Environment vocabulary: alien tendrils, alien tendrils emerging from egg/slime, alien egg cluster, alien hive, biomechanical, infestation, dripping slime
Good at
Stage-controllable transformation (the minor → full progression is a real axis, not just intensity)
Anatomical specificity (corruption stays where you put it)
Egg/tendril environments without a character (use
no humans, sceneryfor still lifes)Stacks cleanly with style LoRAs — tested across multiple style mixes without degradation
Multi-character and pose flexibility — concept doesn't override scene composition
Recommended settings
Weight:
0.8 – 1.0Works at any SDXL resolution; trained at ~1MP bucket sizes
Clip skip: 2 (Illustrious standard)
Best epoch: 25 (clean, no overtrain)
Xenophage is very flexible with style — tested cleanly across multiple style LoRAs without quality degradation or concept bleed. It also stacks well with other adjacent concept LoRAs to push their output toward the xeno aesthetic: try mixing it with venom symbiote, oviposition, tentacle, body horror, or biomechanical LoRAs to give them a corruption / infestation flavor. Recommended weight when stacking with another concept LoRA: 0.3 – 0.8 to let the partner LoRA lead.
FAQ
Comments (10)
Looking very interresting, but with 999 images you can finetune whole checkpoint (if you have free computing time somewhere). Probably could work even better than just lora, because it whould have much more "space" inside of it for concepts.
That's honestly a really interesting prospect! I hadn't ever considered fine-tuning a checkpoint, but hell, with my A100 on Colab, I may as well try!
I was shocked that the image count paired with the 25 epoch training didn't seem to overtune it. Let me know if you have trouble, I have previous epochs available as well if this one is over-tuned to the subject matter.
Seems you've been getting some mileage out of it!
@Shiggles I did. It is not exactly what I expected (for some reason my first impression about it was something like Venom, but without IP ties). But it does look cool.
As about overkoocking - it is indeed overtrained a little bit. Sertain face types, body structure and composition overall are showing more often than others. It is not bad for random pictures, but the more you try to control details, the more you start to fight lora itself. In my understanding it is due to this being not LoRA, but LyCORIS - combination of LoHa (compressed LoRA for generic picture similarity), LoCon (enforcement of general contours, that is marketed as better character reproductability) and LoKr (not having any effect because illustrious is too small).
For generating random good looking pictures it is amazing. For finer details - too big (as expected, there is 2 or 3 separate loras that you can train from your original datase). Good job!
@WannaBeWoman Thank you, and OH MY GOD THANK YOU FOR THE FEEDBACK. I have the dataset organized enough that I can definitely prune down and retrain some sub-models. My testing was pretty limited to the specific images I like to generate, so I didn't notice the areas that it was overtrained.
Funny enough, this wasn't even supposed to be a costume/tf LoRA at first, it started with me just wanting to train an alien looking egg so egg laying didn't always look like someone was squeezing out omelette ingredients lmao. But then my training image prompting kinda slowly drifted and the result was 'eggs as a side aspect of a ridiculously large locon with glowing eyes and tits'.
I think I will move toward splitting off at some point. In the meantime, would a copy of one of the earlier epochs be helpful? Posted is epoch 25, but I have each training epoch and the overtraining may not be as pronounced around epoch 15. It was still producing decent results on that one, they just didn't quite nail the very specific aesthetic I was shooting for yet.
@Shiggles I am not saying it's bad. It is just wery "final", if I can say so. To make a comparessment: you can have a PC prebuilt with optimal specs and price to just turn it on and use (as your lora is set up. To generate beutifull pictures without struggle) or you can have PC scavanged part by part with couple of them spare or wonky or even not intended to be there (as I prefere to act, stacking many small loras, inpainting like hell (34 continuos inpaintings on one latent was my record, lol) and iterating over many questionable results). Both ways are viable and have results.
Speaking about your style, it is indeed remarkable. This is what actually separates just passers by and diligent creators - intent and meticulous appraisal. Though sometimes it is a little bit too much, considering that quite often what you are looking for do exist but in different form. For example there is couple of good loras for egg-laying and speciffic tags for it on booru (egg laying for example, or this lora wich has some bizarre fetish, can be used just for concept, or even this and that, if you can fight off furries from your prompts ^_^).
Having separate small loras with such a good quality whould be great. For example, I havent seen any for not standard eggs. All of them are either bound to some ridgid concepts (pokemon egg actually is the only one and only from couple of loras) or just using existing borring white egg from default chekpoint tags.
Can say same about different types of slimes and tentacles, though there is much more variety.
There is also some very speciffic concepts about symbiotic tf, like facial expressions (LyCORIS whould be actually good there because it much more ridgid), skin texture and "shape" (speaking about occurances, when something is actually moving under the skin or skin itself changing. Good examples is bubbling skin in "Harry Potter and the Hidden Chamber" after potion of transmutation (I'm not native speaker, so I don't know correct names), Mystique transformation, liquid metall from legendary scene in Matrix after a pill and so on) and other details. To speak abouth this lora, some kind of actuall hive/hatchery estetics and props whould be interresting. Some kind of more complex tentacles with purpouse (people on rule34 know about that staff much more) or more stable backgrounds (there is a cool lora for cocoon, but it is stale for now).
P.S. Sorry for my kilometer long blabbering. Lack of normal sociall interractions does scary things with man...
@WannaBeWoman Ha! I know the feeling!
And don't worry, no offense has been taken! I know from playing with it more over the last couple days that you're right in all respects here!
1) It does exactly what it's supposed to do
2) If you ask it to do something besides exactly what it's supposed to do, it'll fight you a bit.
It's definitely the LoRA I've put the most work and learning/research into, and it's definitely pointed me in a couple of great directions for its next generation.
- Separate out concepts
- Add more variety in the training data (or really, REMOVE some of the repetition, as I had a lot of variety in there but over 75% of those thousand plus images is along the lines of 'detailed anime' with similar faces/poses)
And when generating images, I'm actually typically in your same boat! Any concept I'm trying to accomplish, I'm often throwing in 5-10 LoRA's at low weights for everything from styles to specific concepts to backgrounds and layouts. Actually, part of the reason I created this LoRA is because I was using a ridiculous combination of mile-long prompts with strange weights and laundry list of LoRA's too long to keep track of to generate my images, and as I generated the original training data for the egg itself I thought "Hey, why don't I include all of these other things in the training data so I don't have to prompt as hard?"
And here we are! Better than expected, with room to improve, proud of the results but VERY excited for my next burst of ADHD hyperfocus where I plan on experimenting with a lot of your recommendations here.
I have to say, I really appreciate you and all of your effort in providing your thoughts, advice, and feedback! None of it has been taken bad -- I like this cool thing, but you've given me direction to create similar cool things that help accomplish other goals, and I'm excited to do that!
Any plans for an anima version of this?
Not at this time, maybe if I start using Anima. That said, the training data is available if you would like to crate one! All 1000+ images, with tags.
@Shiggles Nice, I'll give that a go at some point.