toyxyzさんのイラストまとめ

@HannahJ4justice @JDespera @icreatelife Random noise is not necessarily used. You can fix the seed of the noise if you want. In this case, the composition of the image is fixed to some extent. It is also possible to visualize the influence of each prompt and design the prompt based on it.

@LAIPAttorney I wonder how AI images generated by methods other than Txt2Img are classified. In the case of this image, it was produced by inputting a hand-made 3D model background, human skeleton, background color image, and text prompt into AI, and combining the result with Photoshop.

Compared to the left side with auto scaling, the right side starts to break the pose faster.

ControlNet openpose bone thickness test #stablediffusion #AIイラスト #pose2image

ControlNet openpose bone thickness test #stablediffusion #AIイラスト #pose2image

When using only Seg, it is difficult to obtain an accurate pose because only the overall mass and position can be specified. If only openpose is used, the rest is random. When using Multi-ControlNet (seg+openpose), the advantages of both models are available.

https://t.co/U92WVOBxiz

@kohya_tech Liono ptimizerをテストしてみました。 Adamで使用した1/3の Lrで似たような結果を示しています。 クオリティの差はまだわかりません。🤔