見出し画像

LECOでSDXLモデルを学習する

とりにく

需要があったので軽くまとめ。LECOとはいったいなんぞや、何ができるんやということに対する解説は以下の記事を参照。とてもわかりやすいです。

では早速学習方法を書いていきます。

環境前提

Python：3.10.8（3.8だと動きません）
CUDA:11.8

環境構築方法（winローカル）

①winローカルの場合、簡単にセットアップできるツールがあるのでそれを使っていきます。

①上記のzipを適当なところで解凍し、install.ps1を右クリックしpowerShellから実行
"Install completed"と表示されたらOK
②examplesフォルダ以下のconfig.yamlとprompts.yamlをコピーして編集。以下、自分の例です。
neg4all_config.yaml

prompts_file: "examples/neg4all_prompts.yaml"

pretrained_model:
  name_or_path: "C:/stable-diffusion-webui/models/Stable-diffusion/XeLblend1.fp16.safetensors" # you can also use .ckpt or .safetensors models
  v2: false # true if model is v2.x
  v_pred: false # true if model uses v-prediction

network:
  type: "lierla"
  rank: 16
  alpha: 1.0

train:
  precision: "bfloat16"
  noise_scheduler: "ddim"
  iterations: 2000
  lr: 1e-4
  optimizer: "lion"
  lr_scheduler: "cosine"

save:
  name: "neg4all"
  path: "./output/neg4all"
  per_steps: 100
  precision: "bfloat16"

logging:
  use_wandb: true
  verbose: true

other:
  use_xformers: true

neg4all_prompts.yaml

- target: "worst quality, low quality, ugly" # what word for erasing the positive concept from
  action: "erase" # erase or enhance
  guidance_scale: 1.0
  resolution: 1024
  dynamic_resolution: true
  batch_size: 1

- target: "poorly drawn hands, poorly drawn feet, poorly drawn face" # what word for erasing the positive concept from
  action: "erase" # erase or enhance
  guidance_scale: 1.0
  resolution: 1024
  dynamic_resolution: true
  batch_size: 1

- target: "mutation, mutated, extra limbs, extra legs, extra arms, disfigured, deformed" # what word for erasing the positive concept from
  action: "erase" # erase or enhance
  guidance_scale: 1.0
  resolution: 1024
  dynamic_resolution: true
  batch_size: 1

- target: "out of frame,body out of frame" # what word for erasing the positive concept from
  action: "erase" # erase or enhance
  guidance_scale: 1.0
  resolution: 1024
  dynamic_resolution: true
  batch_size: 1

- target: "bad art, bad anatomy" # what word for erasing the positive concept from
  action: "erase" # erase or enhance
  guidance_scale: 1.0
  resolution: 1024
  dynamic_resolution: true
  batch_size: 1

- target: "text, watermark" # what word for erasing the positive concept from
  action: "erase" # erase or enhance
  guidance_scale: 1.0
  resolution: 1024
  dynamic_resolution: true
  batch_size: 1

- target: "1girl" # what word for erasing the positive concept from
  positive: "masterpiece, best quality, ultra high res, 8k, clearly fine detailed" # concept to erase
  unconditional: "" # word to take the difference from the positive concept
  neutral: "1girl" # starting point for conditioning the target
  action: "erase" # erase or enhance
  guidance_scale: 1.0
  resolution: 1024
  dynamic_resolution: true
  batch_size: 1

- target: "boy" # what word for erasing the positive concept from
  positive: " masterpiece, best quality, ultra high res, 8k, clearly fine detailed" # concept to erase
  unconditional: "" # word to take the difference from the positive concept
  neutral: "boy" # starting point for conditioning the target
  action: "erase" # erase or enhance
  guidance_scale: 1.0
  resolution: 1024
  dynamic_resolution: true
  batch_size: 1

各パラメータの解説は一番上に張ったリンクが詳しいので割愛。
Promptの設定について解説すると、お約束のnegativeプロンプト概念をモデルから消して、また1girl、boyに含まれる高品質になるprompt概念をモデルから消しています。

③train_leco.ps1を編集
config_fileを作成したものに変更する。

# LoRA train script by @bdsqlsz

$train_mode = "sdxl_lora"

$config_file = "examples/neg4all_config.yaml"

.\venv\Scripts\activate

$Env:HF_HOME = "huggingface"
$Env:XFORMERS_FORCE_DISABLE_TRITON = "1"
$ext_args = [System.Collections.ArrayList]::new()
$laungh_script = "train_lora"

if ($train_mode -ilike "sdxl*"){
  $laungh_script = $laungh_script + "_xl"
}

# run train
python "./$laungh_script.py" `
  --config_file=$config_file

Write-Output "Train finished"
Read-Host | Out-Null ;

④train_leco.ps1を右クリックしてpowerShellから実行
RTX3090で3時間ほど放置して学習終了した気がします（うろ覚え
結果は上のコードでこんな感じ。一応意味はありそう。

以下は無料コラボではシステムメモリ不足で動かなかったコラボ。誰か試してけろ


いいなと思ったら応援しよう！