Yup i fucked up. First try at creating LoRA model and its shit.
I created a custom AI character LoRA using Stable Diffusion 1.5 and kohya_ss on my Windows 11 laptop with an NVIDIA RTX 4050 GPU. I started by collecting 548 images of the subject and used an automated WD14 tagger to generate captions for each image. I organized the images and captions into a structured training dataset and added the trigger word nao_jinguji to help the model associate the images with the character’s identity. I then configured the LoRA training parameters in kohya_ss, including a network rank and alpha of 16, FP16 mixed precision, and the AdamW8bit optimizer. I set a target of 800 training steps and used my GPU to train the LoRA on top of the SD 1.5 base model. Once training was complete, the resulting LoRA could be loaded into ComfyUI and activated with the trigger word nao_jinguji to generate new images of the learned character in different poses, outfits, expressions, and environments.