ver.20260610/20260519/20260425/20260329/20260302
Quality tag&Negative prompt:May or may not be used. / 使っても良いし、使わなくても良いかも。
ver.20260610
暗所が出にくくなっていたのでデータセットを作り直して調整。調整後、darkやnight等の画面を暗くするプロンプトを入れると、黒帯がかなり強めに出るようになってしまったので、それも再調整して是正。
現状、黒帯は全く出ないというわけではないですが、かなり出にくくなったはずです。
クオリティタグないしネガティブプロンプトは必須ではありません。サンプル画像は全てクオリティタグとネガティブプロンプトを付けていません。
※2026.6.10 23:50 サンプル画像を調整前のモデルで作成したもので上げてしまっていたので差し替えました。
Since dark areas were becoming less likely to appear, I rebuilt the dataset and made adjustments. After the adjustments, when I included prompts like “dark” or “night” that darken the screen, black bands started appearing quite prominently, so I readjusted the model to correct that as well.
Currently, black bands aren’t completely absent, but they should be much less likely to appear.
Quality tags and negative prompts are not required. None of the sample images include quality tags or negative prompts.
旧版の概要はdescriptionから各ver.のabout this versionへ移動させました。
Description
based on anima-preview3(https://huggingface.co/circlestone-labs/Anima,https://civitai.red/models/2458426/anima-official).
Special thanks to the authors of the models anima.
ver.20260425
Readjusted based on anima-preview3.
FAQ
Comments (4)
I am a big fan of your Anima finetunes, thank you so much for creating them! <3
Your latest version (20260425) is a big improvement in style and lighting compared to previous versions and I love it. But for some reason, the amount of faulty limbs and fingers increased a lot with this version, as well as some strange repetition artefacts in the backgrounds. I have no experience with finetuning models (only testing them), but have you maybe trained the model with mismatching image sizes?
Up until the previous version, all datasets had a uniform resolution of 1024. However, since the primary resolution of anima3 is now 1024, I created a new dataset with a resolution of 1536 and used it for some parts. In that sense, I think the datasets were inconsistent.
Even setting that aside, perhaps the datasets I prepared simply weren’t good enough.
I plan to rebuild the datasets with a uniform resolution (probably 1536) and try re-adjusting them.
@calculater Nice, please keep up your amazing work! :)
workflow please?