Flux-34551

基于Dev Fp16版本进行了一个优化，和标题一样，使用了34551张图片进行训练。

An optimization was performed based on the Dev Fp16 version. As in the title, 34,551 images were used for training.

真实性和动作对比原版Dev有了一些提升。emmmm，支持nsfw

The authenticity and action have been improved compared to the original Dev. Emmmm, support nsfw

图片大致分为这些比例和数量，The pictures are roughly divided into these proportions and quantities

1024*1024≈5000

1024*1536≈13000

1536*1024≈14000

1728*1152≈1000

1920*1024≈1000

单人女性＞80%，单人男性＞5%，风景和物品及动物＞10%，双人及多人（以婚纱照，写真，旅拍，全家福为主）＞5%

Single female > 80%, single male > 5%, scenery, objects and animals > 10%, couples and multiple people (mainly wedding photos, couple portraits, family portraits) > 5%

亚洲面孔90%，欧美面孔10%，90% Asian faces, 10% European and American faces

我应该没有时间再去更新这个模型了，需要找工作，但可能会融合一下krea，还需要测试，暂时不用期待了。

I probably don't have time to update this model anymore, I need to find a job, but I may integrate it with krea, and it still needs testing, so don't expect anything for now.

Fp8版本已经上传，随便跑了一些图，目前的bug1

男性图出问题的概率增加

bug2 动物毛发一眼假的概率增加。

Version 8 of Fp has been uploaded. I ran some random images. Currently, there's bug 1.

Increased probability of issues with male images.

Bug 2: Increased probability of animal hair looking very fake.

但我没有精力修复了，男性会经常出现购买的一个训练集素材的脸和发型。

But I don't have the energy to fix it, and men often have faces and hairstyles that match the training set materials I purchased.