Version 3
Now i think, that almost 10.000 steps are enough training.
Don't ask me, how long it was. :(
In the next version, I will clean the training images of tattoos and belly piercings, and further optimize the captions to improve compatibility with other LoRAs.
Trained on my PC with One Trainer on Flux1 Dev with 325 images. No AI-generated training images used, all real-life photos. All sample images straight from generation, no inpaint or other postwork.
Hardware is RTX 3060 12 Gb.
Still learning, so dont't beat me. Sample images created with ForgeUI. Mostly of training data images are front view, with some side view examples and few back view.
Works fine with weight 1 and 20-30 steps, Distilled CFG Scale 3.5
Trigger word is "slingshot ??????", but it works with high LoRA weight also without it.
You can also use "classic","tiny", "skimpy", "slutty", "fishnet", "mesh", "seductive", "extreme" in combination with it, to create more "show off" variations. See also sample images prompts for suggestions.
Feel free to try, experiment, cross-post, like, comment, follow and buzz, so i can make this LoRA better in next versions.
Version 2
My second try on LoRA training.
Trained on my PC with One Trainer on Flux1 Dev with 125 images.
Hardware is RTX 3060 12 Gb.
16 hours GPU time. :(
Still learning, so dont't beat me. Sample images created with ForgeUI. Mostly of training data images are front view, with some side view examples and few back view.
Works fine with weight 0.7-1.5, 20-30 steps, Distilled CFG Scale 3.5
Trigger word is "slingshot ??????", but it works with high LoRA weight also without it.
You can also use "tiny", "skimpy", "slutty", "fishnet", "mesh", "seductive" to create more "show off" variations. See also sample images prompts for suggestions.
I think I've tried all the Flux LoRAs for ??????? and ??????? that are available here so far in combination with my LoRA, but the results were not yet satisfactory. The only LoRA that works reasonably well with for visibility is “?????? Pokies v3”
Feel free to try, experiment, cross-post, like, comment, follow and buzz, so i can make this LoRA better in next versions.
Version 1
My very first try on LoRA training.
Trained local with One Trainer on Flux1 Dev with 125 images.
Hardware is RTX 3060 12 Gb.
8 hours GPU time. :(
Still learning, so dont't beat me. Sample images created with ForgeUI. Mostly of training data images are front view, with some side view examples and few back view.
Works fine with weight 0.8-1.5, 20-30 steps, Distilled CFG Scale 3.5
Not tested in combination with other LoRAs, so just try.
Trigger word is "slingshot ??????", but it works with high LoRA weight also without it.
You can also use "tiny", "skimpy", "slutty", "fishnet", "mesh", "seductive" to create more "show off" variations. See also sample images prompts.
Feel free to try, experiment, cross-post, like, comment, follow and buzz, so i can make this LoRA better in next versions.
