lora炼丹术 - 赛博丹炉

学习步数：学习步数越多，细节越多，越多容易过拟合，也就是学过头了。
循环次数：10-20 越多爷会过拟合

 设置

参考

 xyz脚本测试：

xtz plot
prompt S/R    000001,000002,000003,000004,000005,000006,000007,000008,000009,000010,
        0.2,0.4,0.5,0.6,0.7,0.8,0.9,1

        //第一行是将所有模型的000001到0000010测试一遍
        //第二行是不同的lora权重

负面词

 nsfw,NSFW,(NSFW:2),legs apart, paintings, sketches, (worst quality:2), (low quality:2), (normal quality:2), lowres, normal quality, ((monochrome)), ((grayscale)), skin spots, acnes, skin blemishes, age spot, (outdoor:1.6), manboobs, backlight,(ugly:1.331), (duplicate:1.331), (morbid:1.21), (mutilated:1.21), (tranny:1.331), mutated hands, (poorly drawn hands:1.331), blurry, (bad anatomy:1.21), (bad proportions:1.331), extra limbs, (disfigured:1.331), (more than 2 nipples:1.331), (missing arms:1.331), (extra legs:1.331), (fused fingers:1.61051), (too many fingers:1.61051), (unclear eyes:1.331), bad hands, missing fingers, extra digit, (futa:1.1), bad body, NG_DeepNegative_V1_75T, pubic hair, glans, refraction, diffusion, diffraction, nude, open mouth, teeth，(worst quality, low quality:1.4), bad anatomy, low quality lowres, low quality lowres low polygon 3D game, low quality lowres monochrome sketch rough graffiti, low quality lowres very ugly fat obesity scar, low quality lowres chibi, low quality lowres poorly drawn bad anatomy, low quality lowres graffiti unbecoming colorfully, low quality lowres incoherent background, low quality lowres long body, low quality lowres duplicate comparison, low quality lowres sketch retro_artstyle doujinshi, low quality lowres sketch, low quality lowres text font ui error missing digit blurry, low quality lowres JPEG artifacts signature hazy bleary, low quality lowres monochrome parody meme, low quality lowres historical picture, low quality lowres disfigured mutated malformed twisted human body, low quality lowres futanari tranny, low quality lowres tentacle skeleton,watermark, signature, fewer digits, figure, nude, topless, fat,lace,rabbit,big boobs,big breast

 

个人经验

 人物LOSS 0.07-0.09

（1）图像数量先不用太多，不然会消耗更多的时间。

（2）二次元的repeat不需要太多，而实物和人物需要更多细节，需要更高的repeat。

（3）epoch先用10，看一下loss，说是人物loss到0.08左右，二次元0.05左右就可以。从上面有个图也能看出，loss不是越低越好。

（4）unet学习率1e-4。如果loss下不去，就提高学习率；如果loss太低或者没有得到一个特别像的模型，可能是设置高了，再降低些试试。

（5）dim，直接上128就完事。还有个alpha参数，建议是128的一半，不要超过dim。

（6）loss通常0.08左右是最优的，不过也比较玄学，还是用xyz图表测一下。

---

调度器

constat        恒定学习率        （不需要设置余弦）

constat _with_w        带预热的恒定学习率       （不需要设置余弦）

consie        余弦退火       0-1

consie_with_r        余弦退火重启（推荐）每次间隔多少步回到初始值        0-1

linear       线性衰减

BS值低的时候用adam系列

BS值高的时候用lion系列