Next version will be done by Gemini 3. I have manually did 195 images already which took many hours. I just need to train.
This was done with learning rate 0.00007
The reason why I chose this was double-0-seven (007), but with double 0's. It turned out great but I didn't use masking. I learned many things from Gemini 3. This lora took 2 days to train.
Onetrainer:
training resolution: 1328x1328
rank: 16
images: 195
model: float 8bit (w)
Original Creator: https://x.com/vinci194 (Vinci194)
3D Model by: FoxiPaws
Follow the original creator if you enjoy this lora.
They have NSFW in their patreon page here (https://www.patreon.com/VinciTheDragon)
R1.5
Training captions;
overexposed (Comment: Lighting), fishnet lingerie, metallic choker, dragon wings, digital 3d render, large breasts, wearing clothing, ear jewlery (Comment: fuck... its spelt wrong), o-ring collar, arm accessories, smile, angry, very happy, g-string underwear, body chain, bikini, seductive smile, shy, annoyed, grin, big smile, worried, silly face, yawn, sad, exhausted, bikini top, loin cloth, rear loin cloth, v-neck dress, wearing armour and clothing, flower dress, black hoodie, magazine cover, meme format.......Many more, I didn't caption any of the dragon's color.
This lora suffers the same fate as illustrious, it just can't get her likeness correct every generation, mostly her hands, tail, ears, feet.
Description
Qwen2512
AIToolkit
Prodigy
rank 64
training res 1536
Learning rate 1
Why V8? Because that's how many revisions I've done. Its more like 25 but whatever. This was done manually which seems to be a better choice as AI captioning is accurate but it includes false stuff in it and it's too tedious to correct.
I tried fluxb9 and its garbage, ERNIE might be next.