3Dカスタム少女 - 3D Custom Girl LoRA

A LoRA for Stable Diffusion 1.5 checkpoints, trained to simulate characters rendered in the graphic style of 3D Custom Girl, a hentai computer game with a strong emphasis on created characters. Trained on about a hundred or so screengrabs directly from the game.

This new 2.0 update fixes a wide variety of issues the original LoRA faced; as it was my first I had a hard time wrapping my head around how to train it correctly. Fast-forward a few months later, and now I know how to do this shit properly. Main features worth noting include more aesthetically pleasing shots, better prompt adherence, and fixed facial consistency. Good images also don't require extensive ControlNet usage either...

Usage Tips

• The trigger word is "hw3dcustomgirl".

• Most NAI-based checkpoints are generally compatible with this LoRA. Check out the recommended resources to see which ones work the best.

• Use Booru language when prompting.

• LoRA strength should be 0.8 to 1.0.

• Euler/Euler a samplers at 512x768 base resolution. Hires fix is not required.

• Use ADetailer to fix imperfect faces.

• Optionally use certain tags to emphasize the style; try "anime coloring" and/or "cel shading" in your prompt.

Remarks

• This model is heavily biased towards female characters. There are a few male character renders that were featured in the dataset, however please keep in mind that male characters are not able to be made natively in the game.

• Stable Diffusion XL/Pony versions are on my bucket list; I don't have enough VRAM to train these LoRA types yet. Be patient!

• It's ideal to use img2img, inpainting and controlnet techniques to get good images, you won't get perfect images from generations alone.