Experiment with the model and roll the dice a lot!
While you do;
You can buy me a coffee while i train my loras :)
Check for details,
v1.3:
You can use the lora at 0.8 weigth at all times and use the following words to trigger:
" foot, barefoot, bare legs, full body, toes, toenails, perfect feet, black pantyhose, soles "
You can use single or multiple of these words depending on your result. But less is better. Check the images i posted for prompts.
For inpainting:
See this POST. Details in the comment section of each image.
Generate your image and bring it to the inpaint tab. Then, select both feet, select "only masked", randomize the seed, put these in the prompt section AND the LoRa:
You can do this by not deleting all of your prompts and selecting "whole image" but may not work properly.
" toenail polish, (color) nails <feetmasterlora> "
Try and fail between changing the denoising strength starting from 0.4 up to 0.7.
The denosing can solve the toe count, or make it worse. And some colors need extra forcing or adjusting in prompts. You can try following prompting tricks to fix this. Lets say you want BLUE nails:
" toenail polish, blue nails <feetmasterlora> "
" toenail polish, (blue:1.4) nails <feetmasterlora> "
" toenail polish, BREAK (blue:1.4) nails <feetmasterlora> "
" toenail polish, (blue nails:1.4) <feetmasterlora> "
" ((toenail polish, blue nails)) <feetmasterlora> "
" toenail polish, BREAK ((blue nails)) <feetmasterlora> "
Kind of ...
PLUS, i don't know why but choosing "WHITE" has some issues. Either it makes clear nails or if you force too much it breaks the image. Experiment yourself :)
v1.0:
" teasing, 1girl, feet, foot, barefoot, bare legs, toes, toenails, perfect foot, perfect feet "
v1.1:
" teasing, 1girl, feet "
These are the "callers" and if the output is not okay, you can add:
" foot, barefoot, bare legs, toes, toenails "
v1.2:
" teasing, 1girl, feet " - No need to use a word from both sides - " footjob, open toe heels "
" foot, barefoot, bare legs, toes, toenails "
v1.2 Remaster:
This version works with basic prompts. And you can check the trained words from the info in the LoRA details. And you can use these too:
" 1girl, feet, foot, barefoot, bare legs, toes, toenails, perfect feet, footjob, two-footed footjob, open toe heels "
You can experiment with the prompt " feetmaster " too. This word is trained as activation prompt, but not necessary to use. Using this can make the result better or even more mutated. Not tested, not designed to work on activation prompt.
Plus, you don't have to put multiple. Force the ai to draw feet, you can do it with other ways than putting these prompts. Like adding " full body, standing, walking " etc.
Tip: Generate until you get a decent pose, then lock the seed and use hiresfix after. this can save some time.
The toe count can be random sometimes, so you can try to change some values and by exact means " roll the dice " and expect to draw most natural looking feet and toes.
About "Some Values":
The strength of LoRa: Between 0.2 - 0.8 values you can try to get decent results. more strength can break your result depending on your current checkpoint model.
Extra Triggers: If you are getting bad results constantly, check your prompt for extra words that describe what you want. For example, if you put "perfect feet" and "foot" or "feet" in same promps, you may see disoriented, extra spawned or mutated results.
HiresFix Denoising Strength: I highly recommend using HiresFix to get results in similar quality as i shared in page teaser post. Adjusting denoising strength (preferably between 0.3 - 0.6) may fix toe count problem, or worse, add more toes too.
I'll add some examples from my results. Its not perfectly working, but not the worst too.
v1.2 Remaster was a working model of rearranged v1.2 dataset, after many many attempts on v1.3. v1.3 will take a long time to set a better dataset for my plans about this lora.
Please let me know how you're getting results by sharing down below. 👇 🤞
(All the training images are picked from " pixiv, deepbooru, "google search" :P ")