Type | |
Stats | 1,076 5,789 |
Reviews | (151) |
Published | Apr 30, 2024 |
Base Model | |
Training | Steps: 1,210 Epochs: 14 |
Usage Tips | Clip Skip: 2 Strength: 1 |
Trigger Words | v1p3r1seV6 |
Hash | AutoV2 C2A8509727 |
This LoRA was trained on the VIPER RISE (GTB, GT1, V16) game series that I especially like because of the very unique artwork it had. After days of preparation, it's finally done.
Over 90 screencaps of input animations (from GIF sources) were used. Ranging from close up, to mid shots and full body drawings. After that all images were manually upscaled - and then refined (IMG2IMG) using an early version of this LoRA in Krita utilizing the Diffusion plugin.
It was also uncensored this way - and to my surprise this worked almost too good. Image errors (because of upscaling (4x ESRGAN) or poor input quality) were removed with Krita Diffusion, as well. All without straying away from the original artwork.
Tagging was done by hand using booru tags and special refinement on camera positioning, plus additional scene context.
This LoRA does (not) know any character names, but it works REALLY well with other character LoRAs from VIPER - or anything else I experimented with. Please be aware that using too many LoRAs will give you weird / overtrained outputs, so reduce the weights accordingly.
DPM++ 2M Karras and Euler samplers work best. Using the (8-step) Lighting LoRA, I usually go with 14-16 steps with a CFG of 2.5 to 3, Clip 2.