#Made a couple of crazy LORA

1 messages · Page 1 of 1 (latest)

meager scarab
#

Happy Halloween! Almost... Make sure you select the version for your graphics card at the top. FP16 for GTX cards and older and BF16 for RTX and ne...

I found out that my hardware isn't powerful enough to train SDXL LORA, so I compromised and instead upgraded this test LORA into something a little...

#

I'll admit that I made these on NVIDIA/Runpod though. I have an A380, but I use it to enable me to do more while I use AI, like record/stream etc.

coral turtle
#

Nice work!. I like the SD1.5 images that look more like a tree girl.
I have a series I created on retro but modern robot girl that would be fun to train as a LoRA.
@meager scarab How many images did you create to train it?

meager scarab
#

Hey there. Thanks! It was a tough train, moreso than I expected as a first time SDXL trainer, I used 258 4k images, trained at 1536x1536 for the SDXL LoRA. I either trained the same resolution for SD1.5 or 1024x1024.

#

All of the images were generated using my Mile High Style packs to combine the styles of Giger and HP lovecraft and high res fixed using the nickelback upscaler in comfyUI before being re-rendered at low denoise to remove upscale artifacts

#

and then they were used as the training data

#

I started with 1280x720 base model generations, upscaled 3x and then re-rendered at low denoise to produce 4k images

#

All of this on 10GB VRAM