q7-3Dstyle

About this series

This is the third version of my personal training, hence the direct use of "V3" in the version number. Additionally, the concept of "3D" is broad—it can refer to skin textures in 3D animations, figurine models, or 3D games (e.g., those using Unreal Engine 5). This version incorporates data collection across various scenarios, and specialized versions for specific use cases will be released in the future.

About Gameskin-V3

The dataset for this version consists of 70 images, trained at a resolution of 640x640. Strictly speaking, there are still certain flaws: for instance, black-and-white checkered floors are inadvertently included in some samples, the dataset has an overall dark tone, and there is excessive binding to the dark purple skin tones seen in games. I may attempt to fix these issues in future updates.

About Figureskin

The training dataset for this part comes from real-life figurine models, but there's a small issue: they all primarily feature anime-style faces. This isn't to say it's bad, but it lacks detail.

This version has some known issues; initially, it seems to be underfitting and lacks strong generalization ability. I might try to fix this in a future update.

About Someting u want see next lora

Please leave a comment and let me know if you'd prefer to see a new category or an optimization of the existing LoRa platform next.