( Not my Model )

https://civitai.com/models/2755842/anima-qingwei-style?modelVersionId=3116023

( Google Translate ):

Update: The results from the previous version gave me confidence that my experimental methodology was sound.

Therefore, for this official release, I believe I can achieve high-quality style fine-tuning using a formal training set comprising 25,000 images.

This version was trained on 25,000 non-AI-generated images while retaining the aesthetic influence of the 1,200 images used in the previous model.

Compared to the base Anima model, this version exhibits a texture and art style that better aligns with my personal preferences. Additionally, the rendering of fingers and toes is slightly improved—at least according to my observations via X/Y plots. My own testing also confirms that compatibility with LoRAs designed for the base model remains intact; since my trained LoRAs typically use lower ranks, they still work effectively and produce good results.

Old Version: A lightly fine-tuned Anima base model with improved aesthetics. It utilized AI-generated data—specifically images I generated and manually curated to ensure high training set quality.

I also trained a specific tag for my own art style, which can be triggered using the artist tag @qingwei .

Training was conducted solely on the DiT (Diffusion Transformer) components with certain modules frozen; this avoids the loss of core Anima model knowledge often seen in model merging, resulting in nothing more than the intended aesthetic fine-tuning.