What's different about version 2.0?
After over a month of more than 10 attempts, I finally chose to train with DoRA, aiming to minimize its impact on limb positioning while handling larger datasets. Just when my patience was about to run out, I was pleasantly surprised by the results.I welcome everyone to test it out and share their feedback. I also apologize for not being able to keep up with your comments during this time.==============================
This LoRA (Low-Rank Adaptation) is experimental, aimed at exploring the feasibility of training LoRA with Flux.
The LoRA can make the default female appearance in Flux lean more towards an Asian look. However, it has not undergone any facial beautification or "internet celebrity face" training.
Usage:
It is recommended to use weights between 0.4 and 0.65. The latest version of ComfyUI now supports loading Flux LoRA.
Regarding Training:
The training was done using the SimpleTuner script with 2,900 images, taking approximately 20 hours.
Disadvantages/Issues:
There might be slight impacts on the limbs, with a minor decline in the overall performance of the model. More training or other methods may be needed.
It is challenging to fit, and the images may easily collapse during learning.
Testing:
The following are images generated by ComfyUI Flux1.dev-fp8, 25 steps, with prompts and seeds shown on the images:
Description
FAQ
Comments (15)
Amazing. This is the first Flux LoRA that doesn't just slightly change style, or mess up the base model. It works exactly as advertised.
Why is there a noticeable decline in detail and sharpness of the "after" photos, and not just in the face? Is this a consequence of the resolution of the images used for training?
Yes, it's confirmed that this issue exists, but the cause still needs more experimentation. When I trained with the same dataset using the xlabs version of the script, this issue didn't occur. There's still a lot to explore and research in Flux's LoRA training.
male version? just woman?
This isn't an official LoRA, so its performance wasn't my main concern. As an experimental LoRA, it's already done its job. For a real LoRA, I've expanded the biggest dataset and am currently training and experimenting, including male images.
Thank you very much for your work, I like it very much. I have some questions and hope to get your answers: How can I train a lora like this one, so that it does not affect the composition and color of the original image, but only affects its Asian features?
非常感谢您的工作，我非常喜欢它。我有一些疑问，希望能得到您的回答：如何能训练一个像这个lora一样的lora，做到不影响原图的构图、颜色，只影响它的亚洲人特征呢？
我认为最重要的因素是，做好训练集的设计，比如这个lora的目标明确了是只改变脸部特征，那么训练集就应该减少其它内容，并且认真的对训练集进行标注。在这个过程中，我用到了yolo的脸部识别模帮助我处理出需要的数据，然后我编写了一段prompts用gpt api帮助我标注了训练集。
only works on dev-fp8, and produces none-sense on dev-fp16.
I have been using this LoRA fine untill yesterday I started to get images not correctly. Could the ComfyUI update affected this lora?
Can it be used in conjunction with flux1-dev-bnb-nf4-v2.safetensors?
NF4 doesn't support LoRA for now . give a try with GGUF , it should work well
I want Asian milfs though
在NF4上似乎權重要在 0.15~0.25 可以正常工作!
请问一下这个lora需要在哪个底模上运行
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.