Trained 768x768 on 450+ high resolution sharp images of women opening mouth wide, sticking out tongue, showing uvula, displaying teeth, roof of mouth etc. tagged by hand.
Works best with
Lora weight between 0.6 and 0.7,
Low CFGs, between 2,2 and 2,9
Base model Cyberrealistic 4.2 SD 1.5
Resolution 768x768 ,
Performs still reasonable on 768x1024, 1024x768 or even 1024x1024!
(yes that is before upscaling and on sd 1.5)
Trigger/use by following prompt tags:
(tongue out)
(open mouth)
uvula
Example of a prompts that works for me:
DPM++ 2M: <lora:tongue28:0.6215> (tongue out:1.2) (open mouth:1.45) ([uvula::0.07]:0.7)
DPM++ 2M: <lora:tongue28:0.6215> (tongue out:0.7) (open mouth:1.2) (long tongue:0.55)
DPM++ 3M SDE: <lora:tongue28:0.7> (tongue out:1), (open mouth:1.2), uvula, dslr, film grain, Fujifilm XT3
A smaller subset of the training images were tagged with speciality tags.
Use the following in combination with above tag(s) to try to get a 'special' view:
roof of mouth (shows the head tilted upwards with open mouth wide open)
long tongue
smooth tongue (leans towards images with clean even-colored tongues, no lines etc.)
An even smaller subset was tagged with the following tags.
Use these in combination with above tags, and on low weight, as they tend to distort the image quite often when too strong:
throat (shows deep in the open mouth),
closeup (only mouth/tongue in the image, no face or body)
epiglottis (when the view is that deep you can even see into the throat, this is rare, but a few images were present to train on it)
braces (where teeth visible with braces)
Most of the speciality tags I inject just 1 or at most a few steps in the beginning , then remove them, e.g. [uvula::0.07] is only in the prompt the first step of total 40 steps. I find this gives better quality in the end as the model has 39 steps to 'perfect' the image on its own without further lora triggers.
For best results I use following settings:
Sampler: DPM++ 2M, DPM++ 3M (SDE), or DPM++ S2 a
Scheduler: Karras or Align Your Steps
ADetailer for eyes: mediapipe_face_mesh_eyes_only , prompt "<lora:polyhedron_the_eyes:0.7> focus on eyes of a beautiful young girl" with high CFG
Incantations: PAG Scale 0,3, Start 0, End 150;
For complex prompts I add : CFG scheduler Linear Low 0, High 60
Negative prompt: (cartoon) (monochrome) (asian) (overexposed) ((penis)) (out of focus) (blurry) (deformed mouth) (deformed pupils) (disfigured) (blue tongue) (extra limbs) (extra fingers) (Deformed) (plastic skin) (mutilated) (lowres) ((dark)) (boring) (lowpoly) (CG) (3d) (blurry) (duplicate) (watermark) (label) (signature) (frames) (text) (closed mouth:0.6) (eyes closed) (yellow teeth:0.9_
Using dynamic prompt to inject ({a|b|c|d|e|f}{7$$x$$0|1|2|3|4|5|6|7|8|9}:0.8) in prompt to try make face somewhat 'unconventional' to the model (not sure it helps :-)
For upscaling I prefer to use 4x_nmkdSiaxCX_200k , Denoising strength 0.01 (!) , Hires sampler method Euler A on 24 steps.