Recommended prompt/tag = “bl1ndm5k over her/his/their face”
A quick note on using the lora!
SDXL kind of hates taking away eyes in a portrait, and if you watch the lora render, you’ll see they actually will often appear early on and then disappear in later steps. Because of this, the lora does struggle a little bit with some turbo models due to the low step count.
Additionally, this means you will probably get best results by only using the trigger tag “bl1ndm5k” when describing the mask & visor. Using ‘mask’ on its own (or even ‘blindmask’ sometimes!) may remind sdxl that most masks have eye-holes (resulting in things like the mask being moved up to show the eyes or limiting the number of face angles the render can use.)
————————————
Version 2 of this is in the works!!
It will be aimed at improving flexibility, prompt fidelity, and overall quality (the dall-e basic face used in the dataset was part of a test, it is not indented to stay as a permanent feature)
————————————
About this lora and why it is a bit bare bones:
This lora was created partly for fun, but mostly it was created as an experiment in affordable lora training. It is an ongoing personal project of mine to experiment with Tensor's online training in order to find affordable but still effective settings to share with other users.
I am currently working on the first version of my write up of what I have found so far. The draft so far can be viewed here: https://civitai.com/articles/4289/ongoing-experiments-in-training-sdxl-loras-on-*********-part-1
So please let me know any feedback you have on the lora and using it! 😁 Any information or impressions you have to share will help improve both that write up as well as any future lora experiments.
Training comparisons:
Version 1.0 = Prodigy optimizer training
[upcoming versions]
Version 1.1 = AdaFactor training (to be released soon)
Version 1.2 = AdamW8bit training (to be released soon)
