精选225张图片，训练了14400步，使得lora整体更加完美以及接近画师的风格
推荐lora 权重为 (0.25-0.45)
0.25是最低标准，想要接近画风就要0.45，最高0.6，超过0.6就会画面崩坏失真
其他模型我不太清楚，但是部分是可以的，这里推荐coffeensfw_v10，darksushi这种类型风格的，推荐使用二次元风格的模型！
V3和V2版本增添了一些细节，之前V2在出一些图片时候，线条有模糊不够完整，这次是最终版本今后不会更新
在此也感谢各位大家多多支持，你的关注将会是我最大的动力！
在2025年我会关注NVIDIA 50系列显卡，手上的3070已经扛不住了，等50系显卡出就会开始整不一样的。受限于3070 8GB显存，很多训练我都难以执行。例如SDXL，3070训练容易出现报错，显存不够等。
自己的英文自从高考后就再也没有学了，一些英语语法已经退步很多，所以下面用的Google 翻译来翻译成英文的
Selected 225 images and trained 14400 steps to make Lora's overall appearance more perfect and close to the style of an artist
The recommended Lora weight is (0.25-0.45)
0.25 is the minimum standard. To approach the art style, it needs to be 0.45, with a maximum of 0.6. If it exceeds 0.6, the image will collapse and distort
I'm not quite sure about the other models, but some of them are acceptable. I recommend coffee sfw here_ V10, Darksushi style, it is recommended to use the anime style model!
V3 and V2 versions have added some details. Previously, V2 had blurry and incomplete lines when producing some images. This is the final version and will not be updated in the future
Thank you all for your support. Your attention will be my greatest motivation!
In 2025, I will pay attention to the NVIDIA 50 series graphics cards. I can no longer handle the 3070 in my hand, and once the 50 series graphics cards are released, it will start to look different. Due to the limitation of 3070 8GB graphics memory, I find it difficult to execute many trainings. For example, SDXL and 3070 training are prone to errors and insufficient graphics memory.
My English has not been learned since the college entrance examination, and some English grammar has regressed a lot. Therefore, I will use Google Translate to translate it into English
Description
精选225张图片，训练了14400步，使得lora整体更加完美以及接近画师的风格
推荐lora 权重为 (0.25-0.45)
0.25是最低标准，想要接近画风就要0.45，最高0.6，超过0.6就会画面崩坏失真
其他模型我不太清楚，但是部分是可以的，这里推荐coffeensfw_v10，darksushi这种类型风格的，推荐使用二次元风格的模型！
V3和V2版本增添了一些细节，之前V2在出一些图片时候，线条有模糊不够完整，这次是最终版本今后不会更新
在此也感谢各位大家多多支持，你的关注将会是我最大的动力！
在2025年我会关注NVIDIA 50系列显卡，手上的3070已经扛不住了，等50系显卡出就会开始整不一样的。受限于3070 8GB显存，很多训练我都难以执行。例如SDXL，3070训练容易出现报错，显存不够等。
自己的英文自从高考后就再也没有学了，一些英语语法已经退步很多，所以下面用的Google 翻译来翻译成英文的
Selected 225 images and trained 14400 steps to make Lora's overall appearance more perfect and close to the style of an artist
The recommended Lora weight is (0.25-0.45)
0.25 is the minimum standard. To approach the art style, it needs to be 0.45, with a maximum of 0.6. If it exceeds 0.6, the image will collapse and distort
I'm not quite sure about the other models, but some of them are acceptable. I recommend coffee sfw here_ V10, Darksushi style, it is recommended to use the anime style model!
V3 and V2 versions have added some details. Previously, V2 had blurry and incomplete lines when producing some images. This is the final version and will not be updated in the future
Thank you all for your support. Your attention will be my greatest motivation!
In 2025, I will pay attention to the NVIDIA 50 series graphics cards. I can no longer handle the 3070 in my hand, and once the 50 series graphics cards are released, it will start to look different. Due to the limitation of 3070 8GB graphics memory, I find it difficult to execute many trainings. For example, SDXL and 3070 training are prone to errors and insufficient graphics memory.
My English has not been learned since the college entrance examination, and some English grammar has regressed a lot. Therefore, I will use Google Translate to translate it into English
FAQ
Comments (7)
Thank you for your work. Your english is good enough to understand without trouble, thank you for spending the extra time to provide a translation.
Thank you for your hard work
great job, what about the pony xl? I'd love to see a version too. If the problem is that your pc is not enough for lora in xl ver. (Although 8gb of video memory is enough for that if anything) I can use my pc for rendering, of course with the understanding that I need your dataset. And I'll give you the final versions, the settings I know, but you can ask for special ones of course.
I haven't tried XL yet, but recently I've been busy with my studies. When I have time, I'll give it a try
I'm currently working on my graduation project and don't have much time to do XL
:)
@ValosEmed I see, well good luck, xl now has 2 key bases, animage (SDXL) and pony diffusion (PDXL) I would like you to train on the second option. Of course images should be upscaled to 1024 or higher by any means (including upscaling)
@happyhen I have a question, is 8GB of graphics memory really enough？When I use a size of 512 x 768 to create images, black images often appear, and by the time I use them, they already exceeded 8GB of graphics memory
Is SDXL and 8GB graphics memory really enough?
@ValosEmed I make my own loras on 8gb so yeah, that's enough. Although some customization will be required, the resolution of course should be 1024 and the video card can handle that in terms of volume, so you can keep it down. I can give you a mini guide on how to do it, there's nothing complicated just need a couple of fixes, as well as a few console commands to run the training without the interface.