This LoRA is my personal tribute to the excellent FSThicc LoRA by @forsenScoots !
This is NOT an exact conversion of his LoRA as i would rather leave him do it in the future, it is my personal take on this style. The initial dataset was generated using:
The magical Lineweight LoRA by @Bridgewalker
And WAI 17
After that, i made a classical ANIMA training and then fiddled with epochs and layers to get a merge i like (especially, i lowered the impact of the adaln_modulation layers).
There is no trigger words, just add it and enjoy 😊💜
PS: as this is a derivative work of the source LoRA (even if only generated pictures where used), i have set the same "no merge allowed" permission (but as it is known, people don't care 😭).
NB: Due to obvious bias in the training, this LoRA "may" not follow perfectly some prompts.... the dataset is exclusively girls and mostly facing viewer ^^;
Description
Posted on Civitai, upload found elsewhere are FORBIDDEN and done without the author permission.
FAQ
Comments (8)
Idk what to say, I'm very honored!! 😭
Love how this turned out!
Your work is super cool, i hope you'll be able to try your hand at Anima too in the future 💜
fp32? :O
That's a Civitai bug, when uploading LoRA you don't choose the dtype like with checkpoint, but the new UI still wants to display something so, they put fp32. But it's bf16 in reality (at least, that's what it is supposed to be ha ha ha)
@n_Arno Ah, ok :D
I wonder why this is only displayed for some LoRAs. My latest (LoCon) is showing the precision correctly as fp16, but I can't see it on any other of my LoRAs/LoCons. I guess you confused the uploader fiddling with the layers ;)
But still that's a pretty big file size, did you use rank 64 or something or is the file size also wrong?
@RisingV I'll have to check. At one point, i multiplied my bf16 tensors by a python float, but this should not yield a fp32 tensors normally ^^; (It's a 32/16 LoRA)
Annnnd.... you where right. After checking, all layers except the "alpha" ended up being fp32 due to a mistake. I have uploaded the fixed bf16 version which yield almost the same results with half the size 😆
@n_Arno Oh alright, I was wandering whether the layers displayed under tensor card also were false, since they showed f32 as precision, but it seems they were right :D
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.