This is a Castorice character LoRA trained on WAN 2.1 using RTX A6000 on Runpod for one day. The training dataset consists of 50% 3D style and 50% anime style materials, including both images and videos. Videos are essential for adding motion understanding to the model. All materials are tagged separately as "Anime style" and "CG style" to provide different visual effects when specified in prompts.Buy me a coffee <a target="_blank" rel="ugc" href="https://ko-fi.com/dreamweebs">https://ko-fi.com/dreamweebs</a>The trigger word is simply Castorice. Due to the mixed anime-style training data, eyes might appear slightly blurry in some generations - adding "purple eyes" to your prompt helps avoid this issue. All sample videos were generated using FusionX.Regarding WAN 2.2 vs other models: While WAN 2.2 has strong motion understanding, it doesn't perform as well with anime character LoRAs compared to FusionX. The official WAN 2.2 examples focus on realistic humans, and the model excels with real people. However, for anime characters, the official prompt guidelines don't work well - lighting effects become chaotic. For best anime results, I recommend using simple, natural language prompts and avoiding complex lighting descriptions.

Files with SHA256: da2f858d52a1df5b2b612f36f8226dfa53bd8c9e30fa18904fd25e09c5bbaf93