Traied as request. Not very ecessary for SDXL, as it ca already geerate
this type of footwear. However, it aims to provide a cosistetly sharper
shape for poited-toe heels. SDXL V1.0 Based model: SDXL 1.0 Weight: 0.5-1.0, may require higher weight i aime checkpoit, up to 1.3. SD 1.5 It seems like sd always geerates some strappy or platform heels eve though
we put that i the egative prompt. This is a simple Lora that ca potetially
remove the akle strap from the poited-toe heels whe the weight is above
0.4. From what I was testig, 0.4-0.8 seems to give a okay result with the
realistic checkpoit models. SD经常会生成带脚踝绑带的高跟鞋, 用反向提示词限制的效果也不太好, 这个lora就是用来去掉这根绑带的, 且对鞋的形状进行了微调.
使用时权重在0.4-0.8比较好, 更推荐从0.4开始逐步往上提, 0.8以上泛化就不是太好了, 颜色也会逐渐固定成黑色.
用真实系的checkpoit会还原的比较好, 二次元系的也行, 但会有点走样, 且权重得设得比较高才能把绑带去掉. Tips for geeratig leg pose: For simple leg poses, just use prompt ad opepose, this Lora should be able
to geerate that pretty well. For more complicated leg poses, it's too difficult for the curret ge sd,
usually you will get buch of mutat legs ad arms, ad the heels will tur
ito a weird shape. For this problem, you may wat to use edge detectio i
the cotrolet, like cay, softedge ad scribble, the mai idea is to use the
outlie of the picture to keep the overall pose ad heels' shape. First, you eed to have a referece picture of the desired leg pose with the heels o, the throw it ito cotrolet The select the softedge i cotrolet the geerate a cotrolet preview for the picture, ad see if the outlie of the heels has bee detected correctly, if ot, you may wat to use the cay istead. The outlie usig cay is more defied tha softedge. The set the cotrolet's weight to somewhere betwee 0.5-1.0 the write your prompt ad geerate your picture. the lower the weight, the easier to chage the backgroud ad the clothes, the higher the weight, the better the shape of the pose ad heels. If your desired clothes are far differet from the referece picture, you ca use scribble istead, it leaves more space for the AI to draw the cotet, but the shape captured is worse comparig the other two. If all three edge-detectig methods failed, you may wat to modify the cotrolet preview maually to make the outlie more clear ad remove the uecessary cotet. it's pretty simple, just cover the uwated cotet with black colour ad draw the outlie with white colour, eve the pait software itegrated i widows ca do that. 绘制腿部姿势的小技巧: 简单的姿势只需用提示词或者cotrolet里的opepose就可以了, 鞋子基本不会变形 复杂的姿势以目前的SD还达不到可用级别, 经常会生成有多条手臂和多条腿的图, 而且高跟鞋也会跟着变形,
所以我们得用cotrolet里面的边缘检测去限制住大体的姿势不出错. 常用的有cay, softedgehe和scribble,
ormal和dept也能用, 但对高跟鞋的边缘检测没前三种好 首先把穿着高跟鞋的参考图片放到cotrolet里 然后选择softedge作为控制类型, 生成一下预览看看高跟鞋的形状有没有被勾勒出来, 如果没有, 可以尝试用cay, 出来的线条会更清晰 然后设置cotrolet的控制权重为0.5-1.0, 低点儿的话方便换衣服换背景, 高点儿的话形状检测更准确. 设置好了就可以写提示词跑图了 如果你要换的衣服跟参考图片的差距很大, 可以考虑使用scribble, 它能给AI留出更多的创作空间, 但相对的高跟鞋的形状控制就会比另外两种方法差些. 如果三种方法都不能生成你想要的姿势的话, 那大概率是参考图片的背景太复杂了, 你可以手动修改一下cotrolet的预览图, 把无关背景都用黑色盖上, 然后用白色勾勒一下想要的边缘. 十分简单, widows 自带的画图都能搞定 测试下来, softedge对高跟鞋的检测是最好的, 换衣服之类的也很泛用, 所以首先选择softedge来做边缘检测 V2.1 Expad the regularizatio ad traiig set Network dim set to 128 slightly higher prior loss weight with regularizatio 继续扩张训练集以及正则集 网络难度提高到128 正则权重稍微提高了 V2.0 The traiig set was reworked ad expaded, removig all uecessary cotet icludig backgroud ad face, so this versio should work better whe the weight is high. Add dreambooth regularizatio traiig set, the heels' size ad the attach locatio should work better. Recommeded weight: 0.3-0.9, I used 0.5-0.6 i my test image. Work with may clothig Lora whe settig the weight low Not workig very well whe the shoe sole facig up, I thik the problem is with the base model, sice it also ot workig well whe disablig my model 扩张了训练集并且抠掉了背景和脸部这些无关元素, 所以这个版本在高权重下应该不会污染到脸部之类的部位了. 新增了正则化训练集, 现在鞋子的大小和比例以及附着位置应该比之前要好. 推荐权重: 0.3-0.9, 测试图是用0.5-0.6跑的, 推荐逐渐提升权重 在权重较低时能兼容大部分服装Lora 鞋底朝上的姿势大概率会崩, 所以部分躺着的姿势都不太行, 我觉得应该是基模的问题, 因为就算不用我的lora也一样崩. 应该是基模没有对这些姿势进行训练
点击空白处退出提示
评论