Reduce the number of steps, control overfitting, and reduce pollution. Please help with testing. Thank you.
Weight 0.4-1 is acceptableDescription
Reduce the number of steps, control overfitting, and reduce pollution. Please help with testing. Thank you.
Weight 0.4-1 is acceptableFAQ
Comments (29)
Just for my curiosity, How many images in the dataset?
Pose is impossible to change. Enter "(sitting:1.6)", and yet the character still squats 😂
Let me try
@ruanyi ok，晚点再换模型试试
@Ai_00_Girls 隔壁说有很大问题，我都搞不清了，没文字没返图，都不知道问题在哪里。
@Ai_00_Girls 今天我又试了一下，包括，坐，蹲，躺，张开双腿，趴，四肢着地，都能运作，也不用加权重，lora权重是0.7
@ruanyi 我看了你的LoRA metadata，你之后训练别的可以把低模换成1.5的那个底模，或者用NAI也可以。我之前用过一段时间chillout来训练，泛化性挺差的其实，而且像人脸练出来在别的模型用也会变样。还有就是衣服的network_dim其实32甚至16就够了。像那个FreckledVixon做的衣服LoRA都只有9mb应该是用的16或者8，效果也不错。我那个渐变丝袜用的32，不过我是用NAI做低模，也还可以。NAI训练写实图片出来的模型二次元和写实底模都可以用。
@Ai_00_Girls 谢谢，学习了。我下班后回去试试，感谢
No!
This stuff has a big problem. I suggest the download only for testing purpouses.
Soon I will post a sample to show the problem.
Please return to the picture and let me see what the big problem is
@ruanyi Is a problem I never seen before, because it's driving the face to have darker/brown colors.
Is like it was adding a sort of "COCOA" effect on the faces!
The fact you were putting a red background masked the problem!
@ruanyi Sorry for the hurry, CivitAI is full of bug and dont le me post easily. But anyway, with the test I see the background is often trashed and - most of all - the face is completely transformed. These are signs the LORA is driving otehr things, and not only the body, as it should do.
About the quality, you are probably on the right way, since I dont see signs of "stress" on the body and the cloth looks very good.
@settima_ai Thank you for your answer. I should already know what happened, although there have been no issues with the large models I have used so far. Therefore, I have reviewed the marked text and found the problem. There are words describing dark skin in the text, and SD analyzed the jumpsuit as dark skin. I am currently modifying this issue and will release a new version later. Thank you
@settima_ai Hello, maybe you can try a larger model and try again,
@ruanyi What do you mean with "large model"? Tell me an example. please.
@settima_ai Checkpoints
@ruanyi ok, checkpoints. I have hundreds of them. Pirsus Epic Realism, majcMixRealistic, DreamShaper, Deliberate...
What do you mean with large model? 5 Gb instead of 1,99Gb? Or you mean "more popular" ?
Please tell a name.
In any case, the problem is the same for everyone. I think you used 20 images from the same source, perhaps with some color cast. So the LORA learned the dominant color as it would be a common feature to perform. (It's just a theory!)
@settima_ai You can try any checkpoint, I'll try other methods when I get back from work
@ruanyi Hello, sorry for bothering you. But you are experimenting a new way to train LORA and if you find the best solution, then your future LORA will be perfect!
To demonstrate my test is significative, I recall another "bodysuit" LORA I kowk very good for using it since long time and with different checkpoints. Today I made a "face" test with it, and it confirms the quality.
You can see the test here:
https://civitai.com/models/195786/costume-shirt-over-black-or-white-bodystocking?modelVersionId=220107
@settima_ai Hello, thank you very much for your patience. Since I started learning SD in September, you have been the most serious. Haha, tonight I will try to use other bottom molds for training. Once I test it perfectly on my own, I will release it. Please continue to take SD seriously. Thank you
@settima_ai Actually, we can't compare it with this jumpsuit model. It's a clothing set, with leather gloves, leather socks, and a leather bikini. These may all be factors, right
Right, Is clear to me!
@settima_ai I can't find the reason at all. I corrected the words about dark skin in the tag, but it's useless. Now I try to find the problem in the training set album and try again starting from the original image.
@settima_ai When working under Asian checkpoints, the skin color just matches the yellow skin of Asia. However, after using European and American checkpoints, this problem becomes more prominent, and it is very obvious that the face is contaminated.
Hello,
aren't you going to do a version 2 ?
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.