[LoKr + LoRA] Venus Park

Notice:

This model is more for experiments and is not well trained. (it may need extra tags/prompting to reproduce some characterstics)

You should refer to all the example images to check if "the lacking of some features" is due to lacking of some tags.

Usage:

Put this model in to models/lora folder of sd-webui. And use kohaku-xl beta series model to use it.
(This model is trained on sdxl1.0 and I recommend user to use kohaku-xl beta series model)

Introduction

An experiment model for:

• LyCORIS preset config system(using different algo/setting for different modules)

• LyCORIS character model on SDXL arch

• FP8 training for LyCORIS

  • I trained this model with only 8GB vram. (Without cache TE)

• FP8 inference on sd-webui

  • I implemented some FP8 opt in sd-webui test-fp8 branch and all the example images are generated with FP8 UNet/TE

  • Only take 4GB vram when generating the images I post in this model

Training settings:

• Batch size: 1

• Epoch: 50(without any repeat)

• LR: 2e-4/5e-5

• Optimizer: AdamW8bit(weight_decay=0.02, betas=(0.9, 0.95))

• Bucket Reso: 1024x1024(448~2112)

• shuffle caption/caption dropout enabled

More infos

For FP8, I will publish a blog(or note?) for FP8 training and inference recently. Stay tuned.

For preset config system, plz refer to the docs of LyCORIS

For more informations: just comment below.