CivitAi link for download:
https://civitai.com/models/2094601/stabby-toy-sword

(I disguised this LoRA as something meme-ish and harmless to avoid being banned, but the usage guide inside is genuine—please read it carefully for easier control.)

The Illustrious model (and its derivatives) already contains some understanding of the stabbing concept, but it’s abstract and hard to control.
This LoRA is designed to build upon that existing knowledge, making it more precise and controllable through systematic tagging.

Main trigger word: stabbed

Location control triggers:
(Except for “stabbed in the back,” these tags are mutually exclusive.)

abdomen_stabbed

chest_stabbed

head_stabbed

neck_stabbed – weakest tag (limited training data)

stabbed in the back – works only from certain viewing angles

You can also improve accuracy with phrasing like "OOO into XXX",
where OOO is the stabbing object and XXX is the body location.
Example: “long sword into forehead.”

Optional control triggers:

impaling_stabbed – shows both the entry and exit points

assisted_stabbed – another person is holding the weapon

self_stabbed – the stabbed character holds the weapon

independent_stabbed – default; use when you have multiple characters but don’t want any to be the wielder

General prompt template:

1girl [involved character], abdomen_stabbed [stabbed location], impaling_stabbed, [optional control], black closed umbrella [weapon], umbrella into stomach [fine-tuned location], 3/4 view [angle],
[character names, poses, clothing], [background descriptions], [quality/style/artist prompts]


Negative prompt tip:
Adding “multiple weapons” can help reduce duplicated weapons appearing in the image.

Model compatibility & usage notes:
This LoRA works best with anime-style models.
Most testing was done with Wai-Illustrious and Wai-Rouwei checkpoints, using a LoRA weight of 0.5–0.8.
Higher weights increase prompt reliability but can degrade image quality.

For best results, I recommend:

1. Use a higher weight first (to get the composition right).
2. Then use the generated image as Depth ControlNet guidance, and re-generate with a lower LoRA weight (0.1–0.3) for cleaner, higher-quality output.