Aurora正在努力学习如何使用筷子……这个比开炮要难多了。{v1.0}Cough cough, it’s the Chongqing number, I want to make a two-person picture of Chongqing and Yixian hehehe, so I used the official two standing pictures to run out of loraI found that the fitting effect was not good, so I turned to learn LBW (lora layered lora block weight) to remove the style of lora, and I didn’t do anything about the characters and clothing, that is to say, what is it like to restore? ? ?Today, oh, it shouldn’t be yesterday. My 5 well-modified high-definition enlarged frontal images are gone... It made me feel uncomfortable, so I didn’t bother to retouch the images, and directly enlarged them in high-definition and threw them in the post as preview images. Already...you can directly read the prompt words or something.Remember to add "china dress" to the negative prompt words, otherwise there will be a cheongsam. I found that ai can easily take this dress as a cheongsam or something. In fact, it should be a white sleeveless shirt on the upper body and a red skirt on the lower body. matching.trigger: chongqing But I don’t think it’s okay to use prompt words, after all, there are only two pictures.....Also, the weight is about 0.6 to 1. If you don’t like the sharp lines of the picture, you can adjust the CFG to below 7 or lower?咳咳，是重庆号啦，想做重庆与逸仙的双人图嘿嘿嘿，于是用官方的两张立绘跑出来的lora发现拟合效果不好，转而去学了学LBW（lora分层lora block weight）把lora的风格给去掉了，对于人物与服饰没有动手脚，也就是说，比较还原什么的？？？今天，哦不应该是昨天我的5张精心修改的高清放大的正面图没了.....搞得我贼难受了，于是懒得修图了，直接高清放大就丢到post里面当预览图了....可以直接看提示词什么的。负面提示词记得加上”china dress“不然会出旗袍，我发现ai很容易把这个服饰当作是旗袍什么的....实际上应该是上半身是白色无袖衬衣，下半身是红色短裙这样的搭配。trigger：chongqing 不过我觉得不用提示词也行，毕竟已经是只有两张图了.....还有，权重大概在0.6到1左右，如果嫌弃画面的线条太凌厉的话，可以调CFG到7以下或者更低？正面：(((highly detailed, extremely detailed CG unity 8k wallpaper, illustration, highres, absurdres, beautiful detailed eyes, finely detailed breasts, highly detail hair)), ((masterpiece)) outdoors,sky,cloud,grass, ((beautiful detailed eyes)),(Transparent underwear),1girl,blush, pov crotch 1girl,solo,chongqing looking at viewer,smile,one hand holding oil-paper umbrella, bangs,closed mouth,medium breasts, cowboy shot, elbow gloves, red oil-paper umbrella, white gloves,twice red shawl, red skirt,white shirt,shoes, red footwear, double bun, thighhighs, hair bun, white hair, yellow eyes, white thighhighs, hair ornament, braid, braided bun,side slit, pelvic curtain, floral print, &lt;lora:chongqing_1.0_quhuafeng:0.6&gt;负面：(worst quality, low quality:1.4), logo, watermark, artist name ,easynegative,negative_hand-neg,ng_deepnegative_v1_75t,china dress,holding, oil-paper umbrella,{v1.1}I used 27 pictures for training, and finally fixed the clothes and skirts in Chongqing so that they would not become cheongsams...The weight is 0.5 to 0.8. I feel that the hair may be a bit muddy. I will find a way to fix it later. Or you can try to generate a picture multiple times, or use a low redrawing range (0.4 and below) without adding lora after the picture is produced. Running a map to generate a map should improve the picture or somethingAlso, it’s better to increase the resolution. Characters with a small resolution will be very blurry. It takes at least 512x896 to view the image, and then zoom in to 768x1344. The final tiled diffsion should be twice as good.我用了27张图片来训练，终于勉强的固定了重庆的衣服与裙子不至于变成旗袍.....权重0.5到0.8吧，我感觉可能头发有点点糊，之后会想个办法弄一下，或者你可以试试看多次图生图，或者出图后不加lora用低重绘幅度（0.4及以下）跑一次图生图应该能改善画面什么的还有最好把分辨率调大，小分辨率的人物会很糊，大概要至少512x896左右一次出图才能看，再放大到768x1344，最后tiled diffsion两倍应该就比较好了{v1.2}Cough, this 1.2 is actually no different from 1.1, but the number of pictures has been reduced to 20Increase the number of training steps, 1500 steps, 15epoch, 1batchsize, 5repeatThe stability of the clothing is further strengthened, but the opening on the chest may not be so easy to come out (you may need a high weight lora or manually redraw the part by yourself).....At the same time, I recommend using china dress, because the weight of lora around 0.8 can make china dress and china clothes no longer one-piece cheongsam but lora training clothingTherefore, I recommend first using a high-weight lora to produce a low-resolution image, and then use a low-weight image of lora to generate a larger resolution and high-definition zoom (optional)This can improve the reducibility while improving the picture effect and reducing the pollution of lora's painting styleRecommendation weight: 0.5 to 0.8If there is painting style pollution, you can lower the weight of lora, which can reduce the painting style pollution, and I may adjust the painting style pollution in the future :)My lora layer parameters: quhuafeng: 1,1,1,1,1,1,0,0,1,1,1,1,1,1,1,0,0咳咳，这个1.2其实跟1.1没什么区别，只是减少了图片数量变成了20张加大了训练的步数，1500步，15epoch，1batchsize，5repeat对服装的稳定性进一步加强，但胸前的开口可能不是那么容易出来了（你可能要高权重lora或者自己手动局部重绘才行）.....同时，我推荐使用china dress，因为这个lora在0.8附近的权重可以使得china dress与china clothes不再是连体旗袍而是lora训练的服饰因此，我推荐先用高权重lora出一张低分辨率图，再用lora的低权重图生图拉大分辨率与高清放大（可选）这样可以提高还原性的同时改善画面效果并减少lora的画风污染推荐权重：0.5到0.8如果有画风污染，可以把lora权重降低，可以减少画风污染，我可能会在未来再对画风污染进行调整：）我的lora分层参数：quhuafeng:1,1,1,1,1,1,0,0,1,1,1,1,1,1,1,0,0So this is the chopsticks…I`ll be good at it…I guess.

Files with SHA256: d9eb3c31d7f84a91196a37dd1b661ab2e2ff879618d90f5f986fb3e1136b6505