Underboob cloth mod for Stable Diffusion 1.5
Works best for CG\Comic, The class of real photos is prone to NSFW pictures, but most likely only on SD 1.5. Other versions have not been tested, No bias in terms of eyes and face and also not mimicking an individual person,But it affects the size of the breasts as well as the shape of the cloths and hair. So this is the first beta version, comments welcome!
How to use
Simply download and put into your /embeddings folder. Then use 'under-boo' in your prompt. Reduce or increase the weight of the embedding by using this syntax:
(under-boo:1.2) increases the weight, or values below 1 to decrease (such as 0.6-0.8), Too high a weight will affect the overall clothing and hair performance
I recommend putting under-boo to words the end of the prompt, it's strong enough.
Supplement:
If the hairstyle, face shape, and clothing style of your original image have changed a lot after loading this model, you can try to reduce the value of CFG Scale. CFG Scale 5 is recommended
Description
No
FAQ
Comments (7)
I like this TI but it introduces Asian faces. Even with a big negative for asians it still generates them.
Yes, you are right, but it will take a lot of time to optimize this, which is what I am currently short of, so I can't make any more updates to this model, I'm sorry.
Great model, I’m trying to build a similar embedding, but have had very little success.
Could you please share your methodology for creating this textual inversion?
How many images did you use for training?
How did you structure your captions, what information did you include in them?
How many steps did you train for?
What training strength did you use?
https://www.youtube.com/watch?v=2ityl_dNRNw,You can refer to this video tutorial
@franson_liu929 thanks for replying, I’ve actually already watched that tutorial. I haven’t had any problems creating face models. But for whatever reason I can’t get get good results for something like underboob. So I was hoping you could answer some of the questions I asked, in particular your prompt design for the image labels
Well, I probably used more than 200 pictures for mirroring to get more than 400 pictures. The training setting is 20 steps per picture, and the total number of steps is limited to 10,000 steps. The basic model V1-5 of stable diffusion was used for training. output a picture every 500 steps, and automatically save a model every 1500 steps. The editor of the prompt mainly describes the hairstyle, pants, figure, background, and deleted all words about underboob. This is probably the case, but my model was unsuccessful. It was just an exercise during my own learning. This model has too much influence on the appearance and clothing, and there are too many pictures of Asians, and there are too many repeated styles of character poses. Now there are many Someone produced a better lora, so I didn't make further optimizations. hope this helps.
This wrecks their faces, sadly. But it does make underboob.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.