Snatcher (A Hat in Time), Character LoRA by SkittyDitty

Sourced from the Furry Diffusion Discord server. Author notes extracted from the readme, SkittyDitty:

(LoRA trained on terminal SNR FluffyRock e103) Trained at 576px with 39 images for ~2100 steps using Prodigy optimizer

Getting more used to LoRA training, but dataset is much smaller than I'm used to. Strange character, too. Seemed fine after the first and only session I did, so releasing as-is. Head shape in particular seems to vary, but can be adjusted with tuft tags. (Mane and beard tags also seem to cause changes?) Arms can disappear if unprompted for, or by using armless tags. Legs can be generated if desired by prompting accordingly. Seems to handle anthro/humanoid body type okay enough. Dataset has little for behind views, so expect that to be imperfect. Mouth tends to stay open regardless of prompt, but is still possible to close through negatives. (If this is deemed a major issue, I may look into it.) Works fine at a weight of 1, but still has results around 0.8.

Recommended positive keywords: toony, purple body, yellow eyes/mouth, legless, ghost tail, fangs (ghost, spirit depending on how your model handles those tags, or similar)

Recommended negative keywords: naga, white eyes/mouth, hair (flaming tail depending on how your model handles it, or similar tags)

Note: I encountered better results for the head shape with neck tuft in the negatives, which is strange. Mileage may vary? LoRA can do SFW and NSFW alike.