Latest update is the version division between the old artist works and the newer ones.
Pixiv link to original style artist: https://www.pixiv.net/en/users/49675420
Description
SDXL 0.9 version lora trained without TE and pruned to 16 rank
FAQ
Comments (4)
5 More Suggested Resources for SDXL:
[1] First Ever SDXL Training With Kohya LoRA - Stable Diffusion XL Training Will Replace Older Models
[2] ComfyUI Master Tutorial - Stable Diffusion XL (SDXL) - Install On PC, Google Colab (Free) & RunPod
[3] Stable Diffusion XL (SDXL) Locally On Your PC - 8GB VRAM - Easy Tutorial With Automatic Installer
[4] How To Use SDXL On RunPod Tutorial. Auto Installer & Refiner & Amazing Native Diffusers Based Gradio
[5] How To Use Stable Diffusion XL (SDXL 0.9) On Google Colab For Free
dear~ can you share parameters of XL lora train? Especially for the style lora THX~.
Well you can see most of the hyperparams inside the lora metadata. Nothing too extraordinary for an XL, it's just very old known default, just for the test:
AdamW8bit optimizer
Dim=alpha=128 and then resize to 16 dynamical rank after training
bs=4 and gradient checkpointing
1e-4 to ~5e-6 lr range scheduling with slightly edited Cosine Annealing with Warmup(it didn't catched up in metadata somehow, dunno why), but you might use just cosine with restarts with not much of a problem, I hope
Trained during ~400-500 steps per epoch 20 epochs long ~2k steps long totally, divided by bs4 ofcourse, unet only
1024 resolution because of XL
@bakariso thx so much