RHUFSXQ/Personal-Use-MOSS-StyleLoRA

MOSS-StyleLoRA-Personal-Use

A personal-use MOSS-style LoRA for learning and research purposes.

Basic Info

• Purpose: Personal use, learning, and research
• Training framework: KohakuMossTrainer
• Related preprocessing / helper tools:
  • KohakuGBC
  • wd-tagger
  • llama.cpp

Hardware Used

• 64G RAM
• RTX 4070 Ti 12G
• RTX 2080 Ti 11G

Notes

This LoRA is a personal-use model intended mainly for learning, research, and style exploration.
It was trained on / inspired by publicly available works, with the aim of exploring the expression and generalization of the relevant style. No ownership is claimed over the original works, artists, or related styles.

Disclaimer

• This repository is for personal learning, research, and exchange only.
• This model is trained on or inspired by publicly available works. It does not directly copy original works, and no copyright over the original works is claimed.
• Do not use this model for infringement, commercial impersonation, malicious distribution, or any other illegal purposes.
• Generated outputs do not represent the views of the original authors, artists, or any rights holders.
• Users are solely responsible for any risks arising from the use of this model.
• If any rights holder believes that the content of this repository is inappropriate, please contact for removal.

Training Data

number	base model	rank/alpha	lr	batch	steps	images	optimizer	ressolution	precision	trigger	notes
01	MOSS-Image SFT int8	32/32	1e-4	token packing, 6144 tokens/step ≈ 1.5 images/step	1500	~100	AdamW, weight decay 0.01, grad clip 1.0	area ≈ 1 MP, native aspect (mostly 724×1024), no upscale	bf16 compute, int8 frozen base, bf16 latents/text caches	mossstyle

Files and Usage

File	Type	Explain	Recommended Usage
sft-style-lora-01.safetensors	LoRA	Same-batch LoRA adapters	Use the standard LoRA loading method; recommended weight: 1.0; tested and usable without keywords.
sft-style-lycoris-01.safetensors	LyCORIS	Same-batch LyCORIS adapters	Requires a loader that supports LyCORIS; use either this or the LoRA version, not both; recommended weight: 1.0.