Skip to content

Version 0.4.5

Compare
Choose a tag to compare
@kohya-ss kohya-ss released this 07 Mar 10:30
· 1114 commits to main since this release

Updates at 2 Mar. 2023, 2023/3/2:

  • Dependencies are updated, Please upgrade the repo.

  • Add detail dataset config feature by extra config file. Thanks to fur0ut0 for this great contribution!

    • Documentation is here (only in Japanese currently.)
    • Specify .toml file with --dataset_config option.
    • The previous options for dataset can be used as is.
    • There might be a bug due to the large scale of update, please report any problems if you find.
  • Add feature to generate sample images in the middle of training for each training scripts.

    • --sample_every_n_steps and --sample_every_n_epochs options: frequency to generate.
    • --sample_prompts option: the file contains prompts (each line generates one image.)
      • The prompt is subset of gen_img_diffusers.py. The prompt options w, h, d, l, s, n are supported.
    • --sample_sampler option: sampler (scheduler) for generating, such as ddim or k_euler. See help for useable samplers.
  • Add --tokenizer_cache_dir to each training and generation scripts to cache Tokenizer locally from Diffusers.

    • Scripts will support offline training/generation after caching.
  • Support letents upscaling for highres. fix, and VAE batch size in gen_img_diffusers.py (no documentation yet.)

  • ライブラリを更新しました。アップグレードに従って更新してください。

  • 設定ファイルによるデータセット定義機能を追加しました。素晴らしいPRを提供していただいた fur0ut0 氏に感謝します。

    • ドキュメントはこちら
    • --dataset_config オプションで .toml ファイルを指定してください。
    • 今までのオプションはそのまま使えます。
    • 大規模なアップデートのため、もし不具合がありましたらご報告ください。
  • 学習の途中でサンプル画像を生成する機能を各学習スクリプトに追加しました。

    • --sample_every_n_steps--sample_every_n_epochs オプション:生成頻度を指定
    • --sample_prompts オプション:プロンプトを記述したファイルを指定(1行ごとに1枚の画像を生成)
      • プロンプトには gen_img_diffusers.py のプロンプトオプションの一部、 w, h, d, l, s, n が使えます。
    • --sample_sampler オプション:ddim や k_euler などの sampler (scheduler) を指定します。使用できる sampler についてはヘルプをご覧ください。
  • --tokenizer_cache_dir オプションを各学習スクリプトおよび生成スクリプトに追加しました。Diffusers から Tokenizer を取得してきてろーかるに保存します。

    • 一度キャッシュしておくことでオフライン学習、生成ができるかもしれません。
  • gen_img_diffusers.py で highres. fix での letents upscaling と VAE のバッチサイズ指定に対応しました。