Wanted to experiment with making a LoRA for applying textures/patterns to things, so I picked a strong candidate for the most famous texture of all: The Source engine default texture, used when a user is missing a texture in Source engine games. This was most prominently seen in Gmod due to mods using Valve owned textures from Half-Life 2 and Counter-Strike Source that they didn’t redistribute for legal and space reasons. As of July 2025 Gmod contains most of these textures by default, but the missing texture texture can still be seen due to other issues.
Tag: mtex [item] and checkered [item]. e.g., “mtex shirt, checkered shirt”, “mtex socks, checkered socks”, “mtex couch, checkered couch”
To see all my LoRAs and all the example images, please (temporarily) disable X and XXX visibility, otherwise a lot will be hidden at the whims of credit card companies. You can help support making models like this by posting stuff you've made with it to the gallery using the "add post" button, that helps me earn back the training cost at no cost to you.
Made with 21 pics: 7 2D parodies, 8 3D examples (more deliberately added to other games than the actual error), and 6 real items with the texture.
I've seen this kind of thing work by accident (texture of a piece of clothing that wound up applyable to anything) several times so not sure what failed. Wonder if it needs a recognized word to "overwrite"? Maybe it needs a primary article of clothing to apply it to?
V2 works way better. Took till epoch 3 for the couch to get the texture applied, and 5 for the shirt to have the proper checkering. Thus I suspect this kind of trickery (at least in Illustrious) both needs a bit more data than a near minimum size dataset, and has to hijack an existing tag instead of just being able to make arbitrary prefix triggers. Sample pic 3 was a bust since it doesn’t seem to understand this as skin (but there’s none in training data), but started getting generic checkered shirt at epoch 2 and a “no texture” shirt at epoch 5. Not sure if overbaked or Civitai's current issues with drawing magenta are causing issues with it.
Description
FAQ
Comments (6)
Aww snap, I need to uninstall Counter strike source to use this lora! ⌨️
It is working for me, more or less. But you have to add additional tags for making it work. For textiles you have to use print tags like "mtex print" in combination with the clothes you want the print on e.g. "print dress" or "print socks". For skin texture I tried using tattoo or facepaint tags but what seems to work best in the end is using "multicolored skin" and "two-tone skin". Also I had to crank up the weight to 1.5-2.0 in order to make it work.
Thanks. I did get something out of testing at 1.5 power when I initially tested, but it was so awful I didn't post it. Wonder what went wrong for it to be so weak. Wonder if it was a tagging or data failure.
Well, most of the images I generated with it actually suffer from magenta color artifacts, but that's the fault of the civitai generator I guess. I can't say much about this, since I haven't tried to train something like this, but if you'd ask me, I'd have tried to use a lower network dim (rank). I have used network dim 32 only for multi-character LoRas so far. For something like an easy pattern I think I would have tried rank 8 or even lower. The idea here is that with higher rank the model tries to learn additional stuff and the main concept gets watered down. But this is just me guessing. I'd need to experiment myself if this is valid.
@RisingV Was considering lowering it for this, but didn't want to touch too many variables in first attempt. Trying experiments on this has been something I've wanted to do for a while.
Ok, that's certainly one thing you can do. Others would be upping learning rate and total step count obviously. I think learning a pattern may be harder for illustrious than learning a character. However this source missing texture is basically a checkered pattern and illustrious already knows about this. So using "checkered" in the prompt seems to make it work at weight 1.0.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.