仅由一张图片作为原始数据集训练出来的lora，
倾注了我大量心血的女儿
一图流训练的集大成之作
完美泛化+特征保留
技术力的体现！
此lora的训练相当麻烦
使用时权重请设置为0.6
所以如果你喜欢，请给个五星好评，这是我坚持下去的动力！
A lora trained from just one image as the original data set.
My daughter who has poured so much of her heart and soul into
A masterpiece of one-image stream training
Perfect generalisation + feature retention
A demonstration of technical power!
This lora is quite a pain to train
Please set the weight to 0.6 when using
So if you like it, give it a 5 star review, it's what keeps me going!
Description
FAQ
Comments (12)
can you teach how to do that with only one image?
I've done a tutorial like this on bilibili
@Eisthol can you put the link please?
@LadyNuggets Its probably in chinese, i was curious to see their process. Im doing the same thing and it requires a lot more work. For example v1 of my character i started by generating characters with that have the same proportions with short hair color and eye color, then I would photoshop the hair from the good image on it, then i trained a lora. once the hair started looking right I started adding more features, slowly improving my training data. You have to be careful with generating too many of the same type of images. For example, right now my current version has a bias that it wants to put her hands behind her back.
@Eisthol can you provide the link to the tutorial?
@Kamisamatachi @LadyNuggets 【倾囊相授】c站排行前十五炼丹师教你养赛博女儿_哔哩哔哩_bilibili
This tutorial is in Chinese, but you can use the translator
@Eisthol Nice tutorial (the only word I understood was Lora 😂)
How can you train LoRA with only 1 image in data set? I would like to train LoRA of character with 20 images but even that is not enough..
I've trained 2 loras with about 12 images, and it came out really good. Basically it just over trains on the few images it has, so the results can be hit and miss. Once you get a few images the best looking ones you can re-train on the generated images. My next lora will have over 200 images that were generated for the training. It all started from about 12 source images.
I saw a guide on youtube on how it's done, but it's moreso doable with art styles rather than individual characters.
https://www.youtube.com/watch?v=SSbrpAsnK3U
Theoretically it seems really difficult to pull it off with one character, but it seems it is possible after all.
@superskirv this is called loopback training iirc, I've done this method before too. It works well if you're willing to put in the effort to inpaint/draw/correct mistakes on the early outputs.
Which model do you use?
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.