Negative str values make boobs smaller.
Positive str values make boobs bigger.
Likely not as useful for for I2V, since WAN's I2V is first-frame based.
I recommend using both High and Low noise LoRAs for best results. If you only use one, I suggest using the high noise.
On-Site generation is unreliable, as Wan2.2 LoRAs cannot be used with it appropriately.
Description
V1. Trained from multiple LoRAs at Rank 16, merged, then deranked to Rank 4 with minimal quality loss.
FAQ
Comments (19)
This works exactly as described. I am incredibly impressed. Any details with respect to the training process I would love to hear - you clearly had a process with how you approached this.
I ended up modifying the LoRAs directly in python after generating them with Diffusion-Pipe.
The gist:
1. Generate 2 opposing LoRAs (Big Boobs and Small Boobs). The training datasets should be very similar, except for the desired concept.
2. Merge the LoRAs into eachother, with one LoRA at a negative scale. (Big Boobs +1, Small Boobs -1). I calculated the delta tensors for each LoRA, and added them together.
3. Then used singular value decomposition on the merged delta to extract the merged A and B tensor values.
4. I downcasted them to float16, and saved them to file as a safetensor.
The reason why generating 2 opposing LoRA concepts works:
Any unwanted behavior (skin color, sheen, lighting, etc) is present in both the positive and negative LoRAs, so when I do an inverted merge, the unwanted behavior cancels itself out. Additionally, an inversion of "Small boobs" is "Big boobs", so an inverted "Small boob" lora is just a tensor for the opposite, making the desired behavior even stronger.
I also tested another strategy of merging both LoRAs into the main model (again, one inverted), then extracting the LoRA from the model diffs. This allowed me to downcast to a much lower rank (Rank 4) than what I trained the original positive and negative LoRAs at (rank 16).
Since most (not all) of the unwanted behavior is canceled out by an equally trained opposing LoRA, you can crank this LoRA's strength well above 1.0 and still have functioning outputs.
@ComfyTinker Really cool reasoning here. Thanks for sharing.
This is outstanding work and a very very smart way of building this.
Thank you for sharing the method and for the work you put into this.
@ComfyTinker Men will go great lengths, only to generate bigger boba :)
@ComfyTinker Thanks for the info! Do you know if it's a similar concept/training method as this repo? https://github.com/rohitgandikota/sliders
Works perfect! Thank you!
Works in Native, but in Wrapper i got an AssertionError while trying to load
Kijai wrapper? I used the ComfyUI repackage to generate this one. Let me see if I can update some of the key names and re-upload. Stand by
Want to give it another try @JzPz ? I don't have the Kijai models downloaded locally.
@ComfyTinker it loads now and works, thanks.
so this is for t2v and not i2v?
It will work for both. T2V models are more universally usable, but require more careful/curated training to get it right because it also affects the initial image. I2V doesn't affect the initial image, so people don't have to worry about overtraining as much.
Works great with i2v. Works if the breasts are reveal later, example like shirt lift. Needed a lora like this. Thanks!
Nice!
DYING for a 2.2 version!
Hear hear
How do I use this? Do I just load it in as lora and set strength_model to 2 or -1 or whatever?
its a great lora, but if you increase the strength past about 0.4 you start to get these dimples/holes in the nipples
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.