SD2.1 Finetuning

Hi all,

been finetuning SD2.1 for a good while now, it's a very very slow process to ensure it's done right. 

I'm following the same approach I took in the small dataset sample I released a few weeks ago but this time at a larger scale and a slower learning rate. This means it's taking way longer to train but I think it's worth it in terms of quality.


Here's a sample of how it's looking today, still not ready but I'll highlight some pro's and con's so far. Even some of the samples below have issues with anatomy etc so bear in mind this is just a status update and does not reflect how it will look when it's complete

PROS:

-It's taking to BDSM quite well, even things where the ropes are squashing breasts etc it is training quite well

-Versatile datasets, there's multiple types of data going into this, various ethnicities/positions/type of nsfw content such as:
* Cameltoes
* Butt plugs
* Male to Male kissing and penetration
* Female to female kissing and penetration
* Female to Male kissing and penetration
* Oral sex across multiple genders
* BJ
* HJ

  * Bikini tanlines
* Variety of anatomy both male and female / breast / penis sizes + shaved vs trimmed vs hairy, large areolas/medium small areolas, soft / hard nipples etc
* Including Yoga pants datasets
* Flashing breasts and ass datasets
* Penis from multiple angles and types, flacid vs erect vs semi  / veiny + hairy + big balls  + circumcised vs uncircumcised etc
Challenges
- Any type of Penetration is taking longer than any other dataset for AI to learn properly, the current state is somewhat deformed
- Kissing is working but like penetration it's taking longer and it's quite at a stage where it looks normal yet, it's currently showing like to peoples mouths are glued together but in some case it's closer

- Training itself is a challenge, I'm using both Colab and Runpod clouds to train so moving files constantly is a pain and slows it down. I have a GPU being delivered to allow for local training which will be more seamless when used in parallel with one of the clouds (gdrive / colab likely)



ETA

I think 3-5 days of training hopefully if all goes well, then we can invite patreons to test it. It's likely that the release will not be perfect but hopefully it will be a good start to get high rendered quality outputs before scaling on SD2.1, also it will possible serve as a good base model for other groups to train new NSFW data into it.