Emb can replicate every detail of the object, including its body, teeth, cup size and everything else, something that Lora is incapable of doing. Compared to Emb, Lora can only be considered a 'sophisticated wax figure.' Don't assume that Lora, with its hundreds of MB, is necessarily better than Emb just because Emb only has a few KB (I used to think so too). In fact, it's the opposite: Lora is not suitable for replicating objects, as the technical documentation states that Lora has a 'valley of terror' effect when it comes to facial reconstruction, meaning that the result looks more like a wax figure than a real person. Lora is only suitable for low-end computer training and clickbait production shearing the sheep, where high traffic does not necessarily translate into high quality. It's like the difference between a shallow celebrity and a respected seasoned performer;True high-quality models require Emb. Early computers were the size of houses, but now chips are much smaller - size doesn't always equate to quality.The preview images are all REAL, single-generated images, and there is no 'monkey play' image that has been inpainting a second time.The full body facial image is a bit troublesome (generated at once, not partially redrawn). I don't know if it is a defect of sd1.4. My intuition is that sd1.4 is superior to sd1.5 in restoring faces at close and medium distances, but at long distances, 1.5 is better than 1.4. Next, I will make a version based on 1.5 training to verify it.emb可以把对象的一切细节都复刻下来，身体、牙齿、罩杯，一切，这是lora无法做到的。目标左边门牙比右边门牙略短这个细节emb都能捕捉到。lora相比emb最多只能算是【精致的蜡像娃娃】。不要因为emb只有几kb,而Lora有上百Mb就先入为主的觉得lora一定比emb牛逼（我早期也这么想），事实上恰恰相反，lora根本不适合复刻还原物品（原技术文档写明了lora在人脸还原上有一种“似像非像”的“恐怖谷效应”，大白话就是：如同人物的蜡像）。lora只适合低配置电脑体验训练和营销号快速出图割韭菜罢了，流量大不代表品质高，只代表门槛低，如同流量明星和老戏骨的差距，真正的高上限精品模型还得是emb。早期的计算机有房子那么大，现在的芯片才多大？尺寸大小不等于品质的大小预览图没有“二次重绘”的“猴戏图”，都是真实的一次生成图，好的坏的都看的到。全身照的脸部有点麻烦（一次生成，不局部重绘），不晓得是不是sd1.4的缺陷，我直觉sd1.4比sd1.5在中近距离的人脸还原牛逼，但远距离的话1.5比1.4要更好，接下来我会弄一个基于1.5的训练的验证一下

Typical Chinese Beauty Zhang Jun Ning前女友有点像张钧甯

base on sd1.4 full ema 7G, 80 round,next will try base on sd1.5 7G

Files with SHA256: 4c051f7a37a3cbbc972254fa3b0a0175fbfd1f50f44613720e9dfe33b9497c4d