这是一个我的私人模型，采用多个画师的图片训练，并将漆原智志的高光阴影效果进行混合。
能够呈现出非常细腻的细节表现，并拥有非常有趣且惊艳的光影效果。
希望有人能喜欢~
Animagine XL V3版本：https://civarchive.com/models/427631
Pony XL版本：https://civarchive.com/models/430347
Neta Art XL版本：https://civarchive.com/models/460602
SD1.5版本：https://civarchive.com/models/216864
推荐权重：1.0~1.5
推荐模型：FLUX.1
--- --- --- --- --- --- --- ---
his is a personal model of mine, trained using images from multiple artists and mixed with the highlight and shadow effects of Satoshi Urushihara.
It can show very fine details and has very interesting and stunning light and shadow effects.
Hope someone likes it~
Animagine XL V3 version：https://civarchive.com/models/427631
Pony XL version：https://civarchive.com/models/430347
Neta Art XL version：https://civarchive.com/models/460602
SD1.5 version：https://civarchive.com/models/216864
Recommendation Weight：1.0~1.5
推荐模型：FLUX.1
Description
FAQ
Comments (7)
THIS is the quality I want for my Flux Loras. So far it's not good, I don't know what I'm doing wrong :/
Any tips please ? Did you train on Civitai or maybe SimpleTuner, Kohya ? This is looking amazing, althought I didn't know this style for Urushihara, is it based on recent works ?
I'm trying to remake my Lodoss style Lora, after 5 attempts it is nowhere near what you have done here (although the style is different)
这是我想要的 Flux Loras 的质量。到目前为止，效果不太好，我不知道我做错了什么 :/ 请问有什么建议吗？你是在 Civitai、SimpleTuner 还是 Kohya 上训练的？这看起来很棒，尽管我之前不知道 Urushihara 有这种风格，它是基于最近的作品吗？ 我试图重新制作我的 Lodoss 风格的 Lora，但经过 5 次尝试，与你的作品相比差距很大（尽管风格不同）
This is just a test model, it was trained on civitai.
As mentioned in the introduction, this is a hybrid model that combines the styles of multiple artists, including Urushihara Satoshi.
At present, the LORA training of the FLUX model is not very mature. You need to set a higher learning rate and more training materials to achieve a more ideal effect. Of course, too high a learning rate may also cause the screen to collapse, so please try more! Wish you good luck!
这只是一个测试模型，它是在civitai上训练的。
正如介绍中提到的，这是一个混合模型，融合了多位艺术家的风格，包括Urushihara Satoshi。
目前FLUX模型的LORA训练还不是很成熟，需要设置较高的学习率，和更多的训练素材，才能达到比较理想的效果。当然过高的学习率也可能导致画面的崩坏，所以，请多尝试吧！祝你好运！
@flyx3 thanks for answering ^^ (I wasn't sure you could use english so I asked chatGPT to write chinese for me lol)
Well this is strange because I noticed you use 0.0007 for the learning rate, this is slightly less precise then default 0.0005. For example for SDXL I use 0.0003 or 2. I tried your method but I'm not sure how it helped to be honest Flux training is still a mystery ^^ In the end maybe I was expecting too much from Flux, the image is always a bit blurrred with JPG-like noise, then I need to fix it with upscaling to smooth it. I'm also curious about why your alpha/dim ratios are so different, I feel that I still have a lot to learn
谢谢你的回答^^（我不确定你是否会用英语，所以我让ChatGPT帮我写中文，哈哈） 嗯，这有点奇怪，因为我注意到你使用的学习率是0.0007，这比默认的0.0005稍微不那么精确。例如，对于SDXL，我使用的是0.0003或2。我尝试了你的方法，但说实话，我不确定它是如何帮助我的。Flux训练对我来说仍然是个谜^^。最后，也许我对Flux期望太高了，图像总是有点模糊，带有JPG样的噪点，然后我需要通过放大来平滑它。我也很好奇为什么你的alpha/dim比率如此不同，我觉得我还有很多要学的。
@Tetsuoo I am not a model trainer, so I probably don't know what I am talking about.
But one of the big difference between Flux and SDXL is the VAE (Variable Autoencoder) that encodes the image into the latent space, does the diffusion process, then convert the latent image back into pixel space.
Flux uses a 16 channel VAE, whereas SDXL uses a 4ch VAE. This means that Flux encodes a lot more detail.
In practical terms, it means that your training images must be of very high quality. If you have blurry images, then Flux will think that you want blurry images for your LoRA.
So if you have a low quality training set, try to upscale it, denoise it, etc. Make sure it is high quality or your LoRA will not be high quality.
@NowhereManGo @NowhereManGo Thank you for trying to help, this is not easy.
Well, at first I used 1K-2K images from my previous SDXL dataset. But apparently Flux do better with 1024x1024 but with Lora training set to 512... It's all mysterious. So I made all the dataset 1024 square, and that's it. For the pixel quality itself, some (not all) are upscaled if necessary for cleanup the image, then scaled back down to 1024. The result is ok, not great :/
The "Lodoss" style I'm training is old illustrations (1995), very stylized and a bit messy so it's complicate to clean up... The Flux Lora I posted is the 6th or 7th attempt, I got a bit tired after so much testing for DAYS so I will try again but not now, I'm busy with other things x)
@Tetsuoo If the txt2images is blurry, you can increase the steps (for example: 25~30) to generate a normal image.
@flyx3 Thank you, yes it's better around 30 or even 40 steps. But also it's nicer in Forge than in ComfyUI, very strange. With both we get strange grid noise when upscaling, so I will have to work on that too. Upscaling is different with every new model haha