Civitai link: https://civitai.com/models/2431926/stabby-toy-sword-anima

This is just a Beta version for testing training on the Anima[preview1], trained on a pruned dataset of 183 images. A larger dataset will be used once the official Anima base model is released.
The Anima base model itself is already very capable at drawing stabbing scenes with right descriptive prompt (compared to illustrious). This LoRA just give you better controls and refine the details (especially impaling_stabbed and assisting_stabbed).

The prompting guide is written in the civitai post, you can also check the images for references.


>>16035
Dataset with tags and captions(183images): https://drive.google.com/file/d/1Ez3jR7eWMsis45tyAvjNMYJiCLIuJ_vC/view?usp=drive_link
password: gurochan

You can try training this dataset on Chroma or other base model, just have to keep it opensource.
The dataset contains both danbooru tags and natural captions: I vibe-coded a custom pipeline to tag the images first with danbooru tags(using SmilingWolf/wd-vit-tagger-v3), then feed the image along with these tags as grounding into Gemini/Kimi-K2.5 to get descriptive captions in a fixed format (defined through system prompt), then check again manually to fix errors.