
Just dug up the dataset from the depth of my drives and trained this overnight :D. Used 247 in game screenshots to train with AI-toolkit, dim 16 for 6000 steps.
The activation word is "Astarion", but it seems to be associated to the 3d style from the game engine a tad too much for my liking, so I tend to just use "a man with silver curly hair with red eyes". Lora weight of 0.5-0.7 seems to be a good balance for keeping the likeliness while keeping the 3d rendering style at an acceptable level.
Some in game costume works, refer to the prompt in the posted images for trigger words. A higher lora weight would make them more accurate but at the cost of realism to the rendering style from the game.
All showcase images are generated using Forge with Flux GGUF Q8 model & fp16 t5xxl straight from txt2img.
Notes: idk why but cross-post rarely works for the images generated by my Forge, even though I have the Civitai Helper plug-in installed and scanned. The Realism Lora is a nice addition for, well, realism. CCTV Mania is extremely fun to play with. Hope Forge and Civitai could fix the cross-post thingy soon so that I could credit the authors properly :/
描述:
训练词语: Astarion,A man with silver curly hair and red eyes
名称: Astarion-00_edited.safetensors
大小 (KB): 167939
类型: Model
Pickle 扫描结果: Success
Pickle 扫描信息: No Pickle imports
病毒扫描结果: Success