Thank you for your service..
A soldier uniform and outfit LORA for your characters. Trained on mostly realistic images and most of which are NSFW. So this LORA might lean more towards less clothing generation unless prompted otherwise. Also works on female characters.
Weight: 0.65 - 0.85
Trigger: sexysoldier
Additional tags:
Armor: camouflage pants, tank top/t-shirt/shirt/topless male, open clothes, dog tags, gloves, hat/cap/helmet/beret, vest, boots, bandolier,
Weapon: holding gun, assault rifle, machine gun, holster
Note: Some components might not appear properly because the AI thinks the character should be more naked or something. For example when I prompted tank top for my character, he's wearing a tank top but he's also showing his midriff. To avoid that, just use brackets to add more emphasis to the component. In my case, I need to write (tank top).
p/s: If you're leaving a review, consider sharing your generated images too. I love looking at your creations..
Buy me a coffee.. ☕
Description
FAQ
Comments (6)
if it wasn't designed by Hugo Boss, it's not a sexy uniform. nevertheless, a fine lora that will suffice for nearly all, except for those with the most... discriminating tastes
"Dog tags" keep adding dogs to the photos, which is sometimes very cute ;)
Would you be willing to add camo face paint/body paint? I haven't found either a LoRA or a model that can do that (although now I know how to produce clowns).
Also, those guys look damn hot, but "holding gun", "pointing gun", "aiming gun" etc. seems to only work with Clip Skip 1, which makes the rest of the picture not quite there. Clip Skip 2 = the guns sort of hang around… (sometimes literally) Am I doing something wrong?
Will be posting images once I figure out how to do it. Cheers and thanks!
Yeah that's an SD problem rather than a lora problem, it like how 'cowboy shot' keeps adding cowboy hat to the characters. I googled around and still haven't found a decent solution. The best I could find is by using brackets. So put (dog tags) instead..
I never done face/body paint before.. I tried making a body glitter lora once but its a failed attempt. I'll see what I can do but no promises though..
Nah, its another SD problem. The hands are always the trickiest part to generate so making them holding thing like guns and sword are like pulling teeth. It also depends on what model you're using, some are more capable at handing objects in hand. Also realistic models are more 'rigid' compared to 2.5D or anime models since they deal with realism stuff (posing, anatomy, movement, etc) so asking it to make characters hold specific things is slightly more difficult..
Also, try not to overcrowd your prompts with similar tags like "holding gun", "pointing gun", "aiming gun", since it might confuse the AI. Just use one and then try lowering the CFG scale to allow more freedom for the AI to cook.. Good Luck!
@Kairen92 Just an idea, but what about dog_tags and cowboy_shot? or dog-tags and cowboy-shot I personally haven't ever gotten a cowboy hat out of "cowboy shot". If you were to make another version of this particular one in the future, maybe training on "dogtags" would work better than "dog tags". I had a similar problem with the "cocktail" lora (tucked penis, viewed from behind). The creator used "cocktail' as their trigger word, and so I keep getting so many guys drinking martinis with their dicks and asses hanging out, you'd think it was Saturday night at the local drag bar.
@bjornhafthor For the gun problem, it may involve having to generate the main image at clip skip 2 and then inpaint the gun at clip skip 1.
@dita Hmmm, not sure about that.. Because I did hear that SD still autodetects "_" and "-" as spaces so even if you put in "dog-tags" they still detect it as "dog tags". Even putting it as "dogtags" doesn't seem to help. Its also one of the reason why a lot of LORAs uses trigger words with number in them like for example "d0g tags" to avoid SD misinterpreting them as dogs. For some of my models, like this Sexy Soldier outfit for example, I used "sexysoldier" as the trigger because that's what I wanted anyway so its fine if SD interpret it as "sexy soldier".
With the cowboy shot thing, its mostly because it detects the word 'cowboy' so any base model with a lot of data on 'cowboy' looking characters get prioritized, hence the cowboy hats. I have tested this on several models (including mine). Fortunately, it doesn't happen that often and if it did, it only happens like 1 out of 5 gens at most. Although there was that one model (I forgot which), that's heavily biased towards cowboy characters that it generates cowboy hats 2 out of 3 gens..
Btw, most of the things I've said are a mix of personal experience of trial and error and googling around, so it might not be 100% accurate. After all, I'm still figuring out this shit lol. So take it with a grain of salt?
@Kairen92 I have the Embedding Inspector extension installed, so I decided to see how CLIP parses several different tags. As you said, 'dog tags' and 'dogtags' are parsed the same. 'dog_tags' parses a bit differently, but probably not any better. I also tried several other things to see how they parse:
dogtags: dog #4326 tags</w> #11606
dog tags: dog #4326 tags</w> #11606
dog_tags: dog</w> #1929 _</w> #318 tags</w> #11606
dog-tags: dog</w> #1929 -</w> #268 tags</w> #11606
doggytags: do #639 ggy #24935 tags</w> #11606
dog + tags: dog</w> #1929 +</w> #266 tags</w> #11606
dog.tags: dog</w> #1929 .</w> #269 tags</w> #11606
🐶🏷️: 🐶 #14466 \xf0\x9f\x8f #1109 \xb7 #115 ️</w> #1001
puptags: pup #11926 tags</w> #11606
k9tags: k</w> #330 9</w> #280 tags</w> #11606
Even though CLIP parses them into different tokens and all that, it's possible that the U-net could still render those different tokens in a similar way, but probably not exactly the same.
So, to avoid anything that creates a dog inadvertently, you'll want to avoid anything that tokenizes down to include the token for 'dog', which in this investigation is token #4326. You'll also want to avoid any triggers that parse out as 'dog</w>' which means "dog <end of word>" (I think). The trigger 'puptags' would also likely generate puppies from time to time.
Out of the ones I listed above, the safest ones to use might be 'doggytags' and 'k9tags', since CLIP tokenizes them to avoid the tokens for 'dog' and 'pup'. You might be safer with 'dogtags' or 'dog tags' than 'dog-tags' or 'dog_tags' because there's no end of word following 'dog'. That's just a guess on my part though!
The emoji one might work too, but asking users to enter emoji each time they use it is… Well, that's something only I would do. 😆
I'll see your grain of salt and raise you by 2. ;-) In other words, take this for what it's worth, which I admit ain't much. I know how to do what I've shown you, but I can't really explain it beyond what I have. I could be way off with this. But hopefully it gives you some ideas for creating trigger words. You should download the Embedding Inspector extension too (if you're using Automatic1111), to look at how possible trigger words are tokenized when you're making a new LoRA.