New version with all 6 outfits: hoodie, neko, atlantis, kimono, red alter ego, and bikini. As well as a generic "gawr gura" tag for all your other needs.
Observations:
Correction: The activation tag for the hoodie outfit is also "gawr gura".
Use at weight 1, it's slightly overcooked but the face and teeth get fixed with hires fix.
You need very few tags for each outfit to work, check the tags to the right side. But you can still use more tags such as: shark tail, sharp teeth, trident, shoulder bag (neko outfit), black hoodie (red outfit), two side up (default outfit), barefoot/sandals, shark/cube hair ornament, etc.
The bikini is the popular but unofficial shark print bikini, one cup should be red with teeth.
Made using my Lora guide. This model is also available on tensor.art
Description
Works better at 1 weight and has a less forced style. You may still prefer the old v2, try what works best for you. I felt compelled to publish it as my techniques have improved and I used the new min_snr_gamma.
FAQ
Comments (4)
Could you explain better what is this min_snr_gamma thing? I read the repository but didn't understand a single thing 🤦‍♂️
It's a new setting, it treats Lora training as a multi-step process (which it has always been), therefore it adjusts loss over time to make good steps better and bad steps less impactful, resulting in faster and more effective learning. Practically speaking it just makes training more efficient. Right now it's available in my trainer colab and in Lora Easy Training Scripts.
@Hollowstrawberry Got it! I'm running on a 3090, got from 30 minutes to 8 minutes while training, this is wild!
@holostrawberry is it need a fewer training step than before to getting good result?
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.