Sorry for 0.9.21 I've published the wrong epoch.

Tests are currently running, I'll update shortly. Better use v0.8 ;)

No Triggerword, just natural language. Add details to your prompt.
Use "sagging [details][pants], revealing [details][underwear] with [details]"

Example: A photo of a young woman viewed from the front. The woman wears a black blouse and sagging black jeans, revealing white boxers with a green waistband with the yellow text 'I ❤️ SAGGING'.

Example 2: A man with dark-toned skin is standing with his left side facing the camera, wearing red sneakers, a light-blue hoodie, and sagging light-purple sweatpants, revealing black boxers.

For more Examples see uploaded images (look in each version)

It seems to combine well with Briefs and Bulges - FLUX

For the text on the waistband you can use emoticons 😍🍆😈.
You don't want anything to appear on the waistband? Use "with a plain [optional:color] waistband"

Now, dear community, to the details for this model (Details on each version in the release notes for this version on the right):

This is the first LoRA I have trained.

The aim is to be able to generate any person of any gender and ethnicity with any type of pants and underpants. Every pose and every shooting angle should also be possible. And although this already works quite well, there are different difficulties in my learning process in the different versions of the LoRA that I would like to mention.

I currently have 9 versions with 10 epochs each, all of which I am currently testing.

I would like to share the best epoch of each version with you here.

I will also write something about training, tagging, captioning and the remaining bugs for each published version. Maybe this will help beginners like me to avoid these mistakes.

Despite my ignorance and a lot of "trial and error", I can say that the Lora has already delivered quite good results in the first attempts and is probably (still) the best of its kind, at least for the topic it deals with.

Tests for v0.1.1 - v0.1.10 (v0.1.10 published)

v0.2 (too bad to publish)

v0.3 (v0.3.18 published)

v0.4 (v0.4.22 published)

v0.5 (v.0.5.10 published)

v0.6 - not published. As a test, I halved the Unet learning rate... Not a good idea

v0.7 - not puplished. Once again 50% learningrate but rank 8. Probably not usable in this combination and with this number of steps.

v0.8 - (0.8.12 published)

Search AI Models & Addons

very few results