Recommended setting is 1
Description
FAQ
Comments (4)
Hello, could you give me some advice on how to train hair? I'm trying to train Gohan Beast's hair, but it fails a lot when I use it with another style LORA. I'm using the same parameters as this one and another LORA you have for hair, but I keep failing. I'm using OneTrainer because of my limited graphics card.
Hey, sorry for the late response. I used the trainer on CivitAi for my LoRA, so I’m not sure if I can be of much help. But maybe it’s the dataset, how many images did you use for your training?
Also, it depends on which model you used for the training. Did you use IllustriousXL or another one?
@The_LeafMakerGod Thanks for responding. Let me explain the problem a little. I've downloaded several of your loras and tested them, and they work very well. You know how to isolate the concept in a good way, and I can't seem to do that. Let's take your Greek nose LORA as an example. When I use a node(comfyui) that activates the heat map according to the tags that were used, the strength of the tags can be analyzed. As you can see in this image, the “Greek nose” tag is only activated on the face.
https://64.media.tumblr.com/0ad7520797c5fbeccb80ccba71d8abc5/7cb99b001ae5c12d-b9/s1280x1920/3ea556d653a572a0af48a827d94d66255ada83af.jpg
But when I train and use the trigger word from my LORA, not only the hair is activated, but the whole body is activated. My trigger word is “g0h.”
I use a data set of 40 images. That test was with 2400 steps. I use the base “illustriousXL_v01.safetensors.” Remember that I'm trying to train only Gohan Beast's hair.
I tried two datasets, one where only the head and hair were shown, and another where only the hair was shown, blurring the face. This was so that the training would focus only on the hair, but neither worked well. The training tags in my dataset look like this: "g0h, huge hair, huge hair, large volume hair, hair up, pointy hair, pointy hair, upright hair, stiff hair, white hair, silver hair, thick hair strands, single bang, no human face, ears, neck, white background, simple background"
@m16u31
Sorry for the late reply. I think it’s because you’re using the same character for all your images, while my dataset contains multiple different characters with the same features. If the model sees “g0h” associated with a full character silhouette 40 times, it will learn to link g0h = full character, even if the face is blurred.
You could try creating a dataset in Photoshop by cutting the hair from Gohan and placing it on different characters to make the dataset more diverse.
Also, note that the no human face tag applies to the entire image, so you should avoid using it.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.