If you like what I do please consider supporting me on Patreon and contributing your ideas to my future projects!
This is a POV missionary LoRa for SDXL 1.0. It was trained on 1024 x 1024 resolution and works best with a portrait aspect ratio like 1152 x 896.
Ideal weight is usually between 0.8 - 1 depending on the prompt. You might have to go even higher than 1.0 if you're using a longer more complex prompt. It also helps to put the <lora> and trigger at the beginning of the prompt.
Main trigger is: pm1s
Additional tags are: girl, pov, penis
Works best putting them all in the prompt together at the beginning then adding more tags after.
CFG: 7 - 11
Have fun and don't forget to share your lovely creations ;)
Invite to my public discord: https://discord.gg/UFkPTfTDWs
Description
FAQ
Comments (19)
some of those photos are hilarious also in great quality! Loved it!
Thanks!
Really struggling to get this to produce themed output with Text2Image. I see all the examples are inpainting. Please update for better Text2Image results.
works fine for me, I just tested it on another checkpoint too
Any way to make a version of this without the facial expression baked in? I'm trying to get a face with more distress but all I get are bright eyes and open smiling mouths. Otherwise this has promise!
Hey thanks. Your prompt definitely gets closer but the problem unfortunately persists. The mouth just won't close. I used words like "(((clenched teeth))) (((gritting teeth))) closed teeth, (((closed mouth)))" even dropped the Lora to 0.1. No dice, it's locked in there lol.
No matter how you use the word teeth that is going to make teeth show I think so it will have the opposite effect of what you're going for.
It sounds like you are wanting something extremely specific. Like a "distressed faces" lora or something. And I don't do commissions for free and I can't retrain the entire model adjusting the dataset and captions just because you have a thing for girls with closed mouths in distress.
You could also try inpainting the face without even using the lora and seeing if your checkpoint is able to do the kinds of faces you are looking for.
I did another quick test to see what I could come up with:
@Aikagi000 Dropping the weight of the text encoder down to 0 should alleviate some of the baked in facial expression.
<lora:pm1s:0.0:1.0> (The weight of the text encoder is 0.0 and the weight of the U-Net is 1.0)
@adempotent Thats a great tip thanks, I had no idea you could control TE weight within the <LoRA>
@getphat I think you're right. I'm trying what you guys are mentioning but I am just not getting it. like I'm getting close but I'm bending steel at this point. I'll probly have to wait for some kind of expression lora for the look I'm going for =\. Thank you so much for the advice though I'll try to put it to use for other stuff.
I think there's a mistake somewhere. On this page I see "PM1S" in the triggers list, the lora filename is also pm1s but in your examples I see "<lora:pm15:0.8>pm15". I'm not sure the trigger does anything by the way.
I renamed the file to "POV_Missionary_SDXL_v1" because "pm1s.safetensors" really doesn't help for organizing. Imagine if all the Loras had such names, you would need a special tool just to figure out what's in your folders
The reason I use triggers like pm1s is because SDXL doesn't have dreambooth implemented into kohya and its very difficult to train custom tags without it interfering with what checkpoints already know.
The correct trigger is pm1s.
A1111 lets you use preview images on your LoRA files and it works fine for organization.
@getphat A1111 didn't work for me, I got an image terribly destroyed with fake colors, I don't know how to describe that curse lol
It's always VERY slow anyway, something like 6 times slower than Comfy, so... yeah
If you take a look at the images I just posted, I tried : pm15, pm1s, and no trigger at all, they all gave similar result. In A1111 when I call the lora it's named lora:pm15, sorry but this all very confusing
@tetsuoo Disregard any typos on my part its possible I was generating while inebriated
and you are free to use the model any way you wish, if deviating from my recommendation works for you, go for it
@tetsuoo That's you... A1111 works fine for SDXL for me and many others. I've used both it and Comfy and SD.Next and A1111 is still my preferred interface. It's just as fast but there are many variables that this likely depends on.
I like your Lora very much. I can't train a lora model like yours even though I tried again and again . Could you share your train images with me? I don't known how to select train dataset images. I'm looking forward to your reply.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.