Training a character LoRA with kohya sd-scripts: dataset, captions, rank

A LoRA is a small set of low-rank weight changes added to a base model. For a character it is the most reliable way to get the same person in any scene. Tool: https://github.com/kohya-ss/sd-scripts (the kohya_ss GUI wraps it). Dataset - 15-30 images for a character is a common working range. Quality and variety beat quantity. - Variety: different angles, expressions, lighting and backgrounds. If every image has the same background, the LoRA learns the background. - Crop to the base model's resolution family (512-ish for SD 1.5, about 1024 for SDXL). kohya supports aspect-ratio bucketing, so you do not need to make everything square. - No watermarks, no other people. Captions - One .txt file per image, same file name. - Start with a unique trigger word (a made-up token such as "zxchar") and a class word: "zxchar woman, ...". - Caption what is NOT the character (background, clothing you want to change, lighting). What you leave out of the captions gets absorbed into the trigger word. Folder convention kohya reads repeats from the folder name: a folder named 10_zxchar woman means each image is repeated 10 times per epoch. Settings that are reasonable starting points - Network rank (dim) 16-32, alpha half of rank or equal to it. - UNet learning rate around 1e-4 with AdamW8bit; text encoder lower or off. - Total about 1,500-3,000 steps for SD 1.5 characters; save every epoch and compare. Spotting overfitting Test each saved epoch on the same prompt and seed. Overfit LoRAs reproduce training poses, ignore the prompt, and burn colours. Pick the last epoch before that happens, and use the LoRA at weight 0.7-0.9. Background: https://wiki.soapbox.community/wiki/LoRA_and_Fine_Tuning