This is a WAN2.2 I2V HIGH LoRA that is not intended for use with the LOW weights. It is designed to fatten the targeted females in the images. It will fatten the females in the images as your other choice of action occurs. (eating, talking, etc.) It should work with any workflow containing a WAN2.2 I2V 14B model. It has an extensive training set with high-resolution examples and is trained on a detailed natural language set, so it has yet to be determined how far its capabilities can be pushed. If you think it might be possible, then try it! It is known to work on just about any style of image. (It is a high noise LoRA, so it should.)
Description
HIGH model only. Use only images with H & W multiples of 64.
FAQ
Comments (18)
Can I ask why only images with multiples of 64? Just curious
Natural limitation of WAN2.
dont be lazy and add the prompts to your videos dude ,, wtf
Done. Prepare to be underwhelmed. Also, the base images are using older models as high noise LoRAs in effect in an overcomplicated 2 part workflow that I will add.
Yo momma is so fat that this Lora made her thinner
This is unfortunately true. It is a little overbaked, and it will transform blobs back to ussbbws. The next version will be better.
The trigger word for this should be 'Yo mama so fat'
It is all natural language these days. A more continuous distribution of information. Something as discrete as a trigger word is very "Model as Context", which is an AI anti-pattern I'm avoiding in my dataset. That's how this model is able to solve "cartoons, men, and platypuses" despite none of those being in the training set. I taught it the concepts by being explicit in my natural language description about any element or aspect included, which also specifically excludes them from what it's being trained on. That is because the related CLIP embedding is "normalized" (Interesting term, isn't it. It could almost be said that it means removing the unsurprising/uninteresting portion of the information.) out of the context. Accuracy matters, as does a common set of concepts to represent the shared context.
Any chance you can upload dataset too? Would love to make something similar
OP trained his dataset on yo mama's onlyfans
A lora worthy of Dr. Doofenshmirtz "Behold! My Fattenator!"
Hope you caught the Doofenshmirtz video addition to the Fattenator gallery. Thanks for the idea, it was fun, and I had no idea the model would work on men, cartoons, or platypuses.
@PhatButtStudios Haha! Nicely done. And it turned out really well, could've been part of an actual episode. 😆
Would love to see this in a t2v model!
I was thinking a lot of people might. Once I have an improved dataset, I will add that. I missed a lot of opportunities in this first attempt and I have learned the kind of dataset I will need to create a high quality LoRA that focuses on an expanded set of concepts. The text side of it really needs to be refined. This current level of work is fine for I2V, but for T2V, I will really need to up my game. It should work the same for T2V as I2V. The HIGH/LOW separation of WAN 2.2 was badly needed and presents many advantages, but I have the compute to do the 2.1 LoRAs and think that while Phantom still requires 2.1 for the best effect, it is worth doing.
Funniest stuff ive seen lmao
why bro...
Pyrocynical 🚨