沙村広明作品4千枚ほどでFine tuningを施したautismmixSDXL_autismmixPony
出来は良くないが死蔵するのも惜しいので公開
Lorayougazouwoiroirosurusukuriputo
1.画像を最大1024x1024の近似で32の倍数になるように縮小
2.GPT-4-Turboでタグとキャプションの作成
話は変わるがエグい絵が多い画集ひとでなしの恋の画像はプロンプトを拒否されるかと思ったが拒否されたのは4､5枚で済んだ｡エグすぎて何やってるのか理解してない可能性がある
3.VAST.AIでH100を借りて10epoch
This model, named "autismmixSDXL_autismmixPony," was fine-tuned using approximately 4,000 images from works by Hiroyuki Samura. Although the quality may not be perfect, it was deemed too valuable to remain unused and hence has been made public.
Lorayougazouwoiroirosurusukuriputo Script Overview:
Image Resizing: Images were resized to the nearest multiple of 32, with a maximum dimension of 1024x1024 pixels.
Tag and Caption Generation: Tags and captions were generated using GPT-4-Turbo. It's noteworthy that while most images from the intense and graphic collection "Hitodenashi no Koi" were accepted, only four or five were rejected, possibly due to the AI not fully understanding the extreme content.
Training Details: The model was trained using VAST.AI's H100 instance for 10 epochs, facilitating the computation-intensive process.