Copy [costume] into Loader/Mods
add [costume] to your Mods.txt --> point to the [costume] folder CharacterFolders.txt which is already configured.
Note:
Some are good. Most have issues.
Mods are as is. Hope to improve results as automated QA capabilities increase. Some have hairs (some AI, some from community), some have codes. Most don't.
Mostly league characters.
Technical:
Stable Diffusion generated references, sprite swap with a static costume reference for bottoms and tops. Generations use strong ControlNet strength with an SD-Chan reference.
In addition to SCL's sdt-hair lora, these produce (usually within a few pixels) of in-game sprite sizes and shapes. The harder you enforce the pixel boundary, the more character and costume detail you lose. These generations are a trade-off between composition, variety, and automation.
Pipeline: E.g. Few dozen representative images -> Train character LoRA (or download) -> SD generation using WAI-Illustrious base, SCL's sdt-hair LoRA + Character LoRA + ControlNet (0.9, pixel perfect setting) -> // static costume bottom template -> compile (will be entire grey bottoms) // SD generation + standalone utility or Discord SDTSD bot -> swaps sprites w/ SD gen cutouts in place.
add [costume] to your Mods.txt --> point to the [costume] folder CharacterFolders.txt which is already configured.
Note:
Some are good. Most have issues.
- arm cutout on torso
- single arms_front reference.
- thigh,chest,back,backside seam
- breast area missing/too much
- just sucks in general
Mods are as is. Hope to improve results as automated QA capabilities increase. Some have hairs (some AI, some from community), some have codes. Most don't.
Mostly league characters.
Technical:
Stable Diffusion generated references, sprite swap with a static costume reference for bottoms and tops. Generations use strong ControlNet strength with an SD-Chan reference.
In addition to SCL's sdt-hair lora, these produce (usually within a few pixels) of in-game sprite sizes and shapes. The harder you enforce the pixel boundary, the more character and costume detail you lose. These generations are a trade-off between composition, variety, and automation.
Pipeline: E.g. Few dozen representative images -> Train character LoRA (or download) -> SD generation using WAI-Illustrious base, SCL's sdt-hair LoRA + Character LoRA + ControlNet (0.9, pixel perfect setting) -> // static costume bottom template -> compile (will be entire grey bottoms) // SD generation + standalone utility or Discord SDTSD bot -> swaps sprites w/ SD gen cutouts in place.