秋叶1.6版本可以融合XL大模型,关注真实和照片。
WebUi 1.6 version can run the XL merge. this model aim at Realistic and Photo.
Description
Westrealisticv2=westscience-50%(2.5pencil-50%pvc)
这次合成难度仅次于神话V6和ROBOTV2,碰巧做减法得到2D曝光废案(也可能成品亮度过暗),偏一点都是雪花和黑白,然后让真实模型-废案=完全真实。
This time merge is difficult second only to Mythology V6 and ROBOTV2. By chance, the 2D exposure scrap was obtained by subtraction (Completed one maybe dark a bit). If it is a little biased, it will be all snowflake and black and white, and then the real model - scrap = completely real.FAQ
Comments (7)
Very interesting. Feels like if Flux was replicated in SDXL. Nice job!
老哥训练这个模型用了多少图?练了多久啊?
Lora训练过至少要3小时100张,我不是训练师,垃圾电脑是做不了这个活,一个普通MODEL至少要一两个月开机输入海量图片,中间不能中断,普通电脑不可能实现。我的大模型是拼接,用已有的模型按照不同需求能得出自己想要的功能,比如要2.5D模型,需要真实模型+2D,按照一定配比(测试图)来了解模型组合情况,越到后面搭配越复杂。
@JoeLink 感谢分享经验,我最近在给一个公司做一个3500张图的SDXL建筑底模训练,只有一张卡,一看800多个小时,我都怀疑人生了。。。
@update 打标图描述也不一定正确(容易偏离目标),需要一些运气才能完成合成想要的功能,我都是用测试图/需求图,然后能估计出模型特点,再将几个模型融合。
@JoeLink 我用的jc2批量反推,几百张图还好搞,几千张就没办法了,SDXL训练我顶着23.7G显存爆的边缘去跑,不敢开训练图像预览,没事,再忍忍就快过去了
处理一下V2画面过暗像素过低问题,以前1.5的经验不是很管用,一般来说是用2D模型再回填一下...
Deal with the problem of V2 screen being too dark and too low in pixels. The previous experience of 1.5 is not very useful. Normally, less than 20% 2D model will be backfilled...















