I trained an Aesthetic Anime Style LoRA for anima p3 using 20,000 highly curated anime images.

A creator released an aesthetic anime-style LoRA (Low-Rank Adaptation) for the Anima P3 image generation model, trained on 20,000 carefully curated anime images. The LoRA is designed to raise the baseline quality (the floor) of outputs rather than pushing the absolute maximum potential. It suppresses overly vivid or flat results and guides images toward a cohesive aesthetic with adjusted spatial lighting and tones. The effect is most noticeable when using short prompts or no style tags; users already achieving great results from the base model will see only subtle support. The creator also included a secondary LoRA called 'sdxl_glossy_lora', trained on 1,250 glossy SDXL images, which replicates the typical highly glossy AI look.

Training was done using the open-source Anima-Standalone-Trainer on Windows with 16GB VRAM (possibly 12GB). Primary LoRA settings: 1024px resolution, learning rate 1e-4, AdamW optimizer, rank 64, batch size 4, gradient accumulation 16 (effective batch 64), and 15,500 steps (~48 epochs). The creator noted a slower-than-expected training and suggests 2e-4 lr might be better. The glossy LoRA used rank 32, effective batch 32, and trained faster. Full inference workflow and model weights are available on Civitai.