Hi there, this model is intended to help when generating images of a tag that was not properly understood by illustriousXL ("nose_picking").
There was a lora that tried to train the same concept but it was for SDXL and it had poor results on Illustrious version, so i decided to step in and train one by my self (holy moly, you have no idea on how much i wanted to drop this project).
Before discussing the versions, i'd like to summarize the lora in few points.
Does it works? Yes, it does
Is it stable? Yes, it requires minimal tweaks to achieve good results
Why should i use this and not other ones? Because (as i am writing) there are no alternatives.
How can i achieve results like the ones in the preview and in the version section? I'll try to write an extensive "tips" section trying to resume all the knowledge i have about this lora, but you can always use the gallery images as reference (i'll publish a few more tomorrow).
I like the model, how can i help? If you like my work, that required way more than i am willing to admit, just leave a review, upload one of your creation and share it with your friends.
The majority of usage tips are available in the version section, but if you are curious about more tricks, tips and info you can read the section below this one.
This version is intended to help in the generation of images that use the tag "nose picking", sadly illustriousXL has not a great understanding of that tag, and often produce images of other gestures (the "finger_in_mouth" one is pretty common).
The v2 is (FINALLY) able to fix this issue, and even toh it didn't turn out of i had envisioned it, it works pretty good.
The original plan was to have a plug and play lora with the activation keyword "nos3pick"... But the first attempt training this the lora i had was kinda glunky to use. (it still worked better than the SDXL one) , so i generated a second dataset by cherry picking the best results of my tests, and then attempted a retrain, and this is the result.
The lora is pretty much plug and play and you will achieve an image like this:
Ofcourse you will be able to change the expression as you like but you might need the additional keyword "nose pick" to make sure the model gets what you are trying to achieve.
As the last time i will show you the difference between not using the lora and using the lora:
Not using the lora vs Using the lora.
As you can see the lora effectively does its work (and because the image was simple enough i was able to just plug it and use the activation tag without any further adjustment).
Most of the test were done using Hassaku (Illustrious) and WAI-NSFW-illustrious-SDXL.
The dataset used to train this version had 100 different images and it was trained for about 13 epochs with 2 repeats. (Yes those are a lot of steps, but trust me they were needed.)
The settings that i often used were:
Sampler: Euler A
Steps: ~28 ~30
CFG: ~6
The images were upscaled and treated with adetailer just to achieve an even better final result.
Upscaler: 4x_NMKD-Siax_200k
steps:15
denoising:0.3
Adetailer:
face
fulleyesdetection
Note: adetailer was NOT USED to help hands in anyway <- So the gesture results are 100% fruit of the lora
(even toh i have to mention that in few limited cases upscaling can fix images with the finger not entirely in the nose or if the nose has not generated correctly).
The prompt structure i followed was:
{Your prompt}, <lora:Nos3pick:1> nos3pick, nose picking
sometimes you might want to add the keyword "nose picking" to further stabilize the output, or if the image is too complex (it is also required when trying to make some complex expressions).
So how do you use this:
activation keywords: nos3pick,
[kinda optional]: nose picking
lora weight: ~1 (1 is the optimal value)
Generally speaking i often stay at 28 steps and 6 cfg, and then try the following:
Generate the image with just the activation keyword, to see if image needed the second activation tag, most of the time the image was good without the optional tag, but when it wasn't i'd attempt to add the second tag (it usually fix everything), and if the seed is particularly bad and even this doesn't solve the issue i'd try to generate activating the extra, with a little of seed variation.
(note: this last case happened just in 1 case, and the image i had to achieve was a totally chaotic in the prompt)
Did i test it enough? Well, let's say that my gpu is ready to be promoted to a toaster.
The first test was run at a resolution of 832*1216:
(the tests for each resolution are 2 couples of before/after images)
Note: the only things between the two proposed images is the presence of the lora and the two activation keyword.
The second test (896*1152):
The third test (768*1344):
The last and most important test was the one related to known characters, and to test it i've run few generations of images using characters from my character's lora to test the compatibility:
Even toh it is not required (most of the time), the tag "nose_picking" has a great stabilizing effect on images and negative effect in using it in the positive after the "nos3pick" tag. So use both of the tags if you can.
The standard expression is pretty bland, but when using both tags, you can freely personalize it without worrying.
This lora doesn't require adeiler, but i strongly suggest the use of it to further enhance the output image quality.
Tags such as:
open mouth,
open smile,
clenched teeth,
half-closed eyes,
glasgow smile,
smirk,
;d, upper teeth only
etc etc.
are fully supported.
Here are some additional things to keep in mind:
If you don't explicit indoors/outdoors and if you don't describe any background, the lora will probably produce an indoor image.
It is highly suggested to not use the keyword "portrait" in the positive prompt.
it is suggested the use of the keyword "close-up" in the negative.
The lora is trained on a dataset of "woman" images, so you will need to lower the weight of the lora a lot to use it for men (to use it on brook i needed to lower it to 0.6)
I hope this lora will be of help for someone of you, i doubt i will ever touch this gesture ever again so i can state that this is the definitive version of it (atleast for the near future).
Thank you so much for reading all of this (i know i talk too much); hope you enjoy this lora, and to see you for my next models.
activation keyword:
nos3pick, nose picking
Weight: ~1
CFG: ~6
It's suggested the use of an upscaler.
1. The rights to reposted models belong to original creators.
2. Original creators should contact SeaArt.AI staff through official channels to claim their models. We are committed to protecting every creator's rights.