QuakeSkin Diffusion

QuakeSkin Diffusion is trained on UV unwrapped texture layouts for the Quake 1 Ranger model and can generate a wide variety of novel character types not seen in the small dataset (12 random skins from https://www.moddb.com/games/quake/addons/quake-1-skins-pack). The training was run until the model learned the concept without reproducing any of the dataset examples.

Instructions

https://www.dropbox.com/s/3p06uurglozq8ya/QuakeSkin_Input.png?dl=0

To ensure the correct layout use the QuakeSkin_Input.png in img2img at 512x512 resolution. In my testing I used euler_a sampler @ 50 steps and CFG 12 with a Denoising Strength of 0.75.

Prompt your subject followed by trigger words in demoura artstyle

The front and back usually match up but not every time so generate in batches to pick the best ones.

To finish the skin you can feed the result back into img2img at 1024x1024 resolution, using similar or more steps + CFG, with a Denoising Strength of 0.2 - 0.3 (as sometimes it might lose the back of the head and do two front faces)

To view your skins correctly on a player model, find and download Ranger (A-posed & Rigged) model and scale the UV map in the Y dimension until the textures fit properly