此lora推荐使用turbo lora 1536*1536分辨率,它与正常工作流的出图效果相差不大,并且速度显著增加.在1024*1024场景下,眼睛细节容易模糊.图片里有我的turbo lora双采工作流,可以自行查看.

仅推荐使用anima base v1.0.其余微调请自行测试

(正常工作流的1024*1024和turbo lora工作流1536*1536的速度差不多)

anima的质量词系统实在太烂了,所以我推荐你不论什么情况都不要在正向提示词使用score评分系统,这会对你的图片造成未知影响.

上一个版本实在太烂了,所以我把他删了

这一版本优化了训练集,将原本在wiki上下载的低清图片换成了300张官方游戏解包的高清素材,由人工手动裁切.

此版本的素材集包含部分全景图片,所以背景的表现也相当不错

This LoRA is recommended to be used with the turbo LoRA at 1536×1536 resolution. The output quality is comparable to the normal workflow, but the speed is significantly faster. At 1024×1024 resolution, eye details tend to become blurry.The accompanying image shows my turbo LoRA dual-sampling workflow — feel free to take a look.

It is only recommended to use Anima Base v1.0. For other fine-tuned models, please test on your own.

(The speed of the normal workflow at 1024×1024 is roughly the same as the turbo LoRA workflow at 1536×1536.)

Anima’s quality tagging system is really terrible, so I recommend that under no circumstances should you use the score rating system in the positive prompt, as it will have unpredictable effects on your images.

The previous version was so awful that I deleted it.

This version has optimized the training set, replacing the low-resolution images originally downloaded from the wiki with 300 high-resolution assets unpacked from the official game, manually cropped by hand.

This version's training set also includes some full-view images, so the background rendering is quite decent as well.