Training on "black sclera".
I have been working to integrate various unique features into LoRa's (and now LoCon). I have found making individual trainings on each feature, then merging the ones needed together gets remarkable results.
Eyes are hard to get right, this helps but isn't a perfect solution yet though.
This has been trained on about 80 AI generated images from various models (picked out from hundreds), after using a couple not so great lora's I had experimented with before.
This can be used stand-alone to add black sclera with much higher efficiency, or mixed with other LoRA / LoCon to mix the feature in. I recommend to avoid using this with multiple other LoRA / LoCon without mixing first, it is more prone to cause issues.
If you like this, get me a coffee :D
Description
Increased the number of tags and lowered the dim for the training.
The results are much, much better.
The bias does not seem nearly as bad anymore.
Improvements can still be made, and I intend to continue to improve it.
Any feedback is appreciated.
FAQ
Comments (5)
AAHHH you beat me to it I was thinking of this stop reading my mind DOWNLOADS HUGS AND STEALS* XD
I can't get this to work at all. I already have the extension installed, yet it says "IndexError: tuple index out of range". Any help would be greatly appreciated. I feel like I should specify that I have tried lots of different models.
which version? I have not run into that, but it might be vae related. Did you upgrade to torch 2? or use any special command line arguments?
@felor I have tried using both verison 2 and 3. Neither worked. I have tried all sorts of vaes, and even no vae. The only command line argument that I have is --no-half. I have not upgraded to to torch 2.
Not sure what it could be, what kind of GPU? AMD has some issues still compared to nvidia, but of others are working this should. I did read there was an issue with a recent A1111 update, that could be it as well.