Sign In

SDXL LoRA train(8GB) and Checkpoint finetune(16GB)

129
1.5k
20
Updated: Nov 23, 2023
script
Type
Other
Stats
344
Reviews
Published
Nov 23, 2023
Base Model
SDXL 1.0
Hash
AutoV2
44EF4A6C58
default creator card background decoration
bdsqlsz's Avatar
bdsqlsz

Credit:

This script package from bdsqlsz

use kohya-ss/sd-scripts (github.com) for core

original script from Akegarasu/lora-scripts: LoRA training scripts use kohya-ss's trainer, for diffusion model. (github.com)

How to use:

0、(windows)Give unrestricted script access to powershell so venv can work:

  • Open an administrator powershell window

  • Type Set-ExecutionPolicy Unrestricted and answer A

  • Close admin powershell window

1、Unzip this to anyway you want(Recommend with other train program which has venv)

if you Update it,just Rerun install-cn-qinglong.ps1.then enter N

2、Run install-cn-qinglong.ps1 in windows(linux just use commnd line)

it will automatically install environment(if you has venv,just put to over it)

3、Put your datesets in /input dir.

a formattting:
./input/2_lora/(datasets images)

4、Run tagger.ps1 for tagger them,automatically.(linux use tagger.sh)

5、Edit your datesets if your want

6、Edit train script(train_8Glora or train_16G_DB)(linux use train_tacanime)

just for

  1. train_mode(lora、db、sdxl_lora、sdxl_db、cn3l)

  2. pretrained_model

  3. train_data_dir

    a formatting ./input/

  4. resolution(SDXL for 1024*1024,sd1.5 for 576*576)

  5. max and min bucket repo(SDXL 640~1536,sd1.5:256~1024)

  6. optimizer(SDXL:adaFactor for batch size1,Pagedadamw8bit for batch size 4)

other dont need more change if you dont know it

if i can help you ,i will very happy that~

Support 青龍聖者@bdsqlsz on Ko-fi! ❤️. ko-fi.com/bdsqlsz - Ko-fi ❤️ Where creators get support from fans through donations, memberships, shop sales and more! The original 'Buy Me a Coffee' Page.