Comment by @voiceofket • Hey

Powerful images. Would love to hear more detail on how the AI was trained.

Stats

Comments

There are two models. One SD 1.5 and one SDXL. The first one was trained with dreambooth. I think it was the LastBen GitHub. The SDXL (full checkpoint) was trained with kohya-ss. Additionally there are two Loras. Each one trained for the Main characters Elliott Hove and Rila Fukushima. For sone recipes we add the Loras sometimes but never fully to create somehow imperfect lookalikes of the actors and not clones training set was just around 25 images for the checkpoints. So the model doesn’t know much about the film but got only a glimpse of the style and characters and imagines a lot. I’m definitely planning to make a video about the whole art generation system.
Thanks 😍 tagging @lens/simonj