The model was trained on approximately 1600 images of various types. To ensure versatility, it was trained on multiple styles, and secondary images were assigned lower probabilities.
Approximately 25% of the main data consisted of fursuits. The rest were images of various different styles.
NSFW training was conducted, but the interweaving of different concepts and the varying learning difficulty make it unreliable compared to dedicated models.
Due to the limitations of the device's memory, the training resolution was set at 512.
The intensity can range from approximately 0.2 to 1.5.
Some display images:
Sampler
er_sdr can obtain more detailed images.
euler/a is more stable.
If you want to obtain better details/quality, use multi-level sampling and switch the sampler.
Training tool: OneTrainer
Rank: 48
If you wish to continue training this model with new data, this setup should allow you to do so; other parameters should not hinder the training process.
If you need better, smoother fur textures, you can use the raw model.
如果你需要更好更细腻的皮毛纹理,可以使用raw模型
====
模型在约1600张多种类型的图像上训练，为了泛用性在多种风格上训练，但次要图像使用更低的概率。
主要数据中大约25%是fursuit。其它是各种不同风格的图像。
进行了NSFW训练，但是不同概念之间相互交错和学习难度不同，与专用模型相比并不可靠。
受限于设备显存，训练分辨率为512。
强度可以在大约0.2~1.5生效。
采样:
er_sdr 获得更多细节
euler/a 更稳定
如果你想获得更好的细节/质量,使用多级采样和切换采样器.
训练工具:OneTrainer
rank:48
如果你想在此模型上使用新数据继续训练,这应当能能让你继续,其它参数应该不会阻止训练
Description
This is an experimental version model trained on the AI-Toolkit, with a relatively small amount of data used and less training conducted.
Due to the poor usability of AI-Toolkit, I later switched to OneTrainer.
===
这是在AI-Toolkit上训练的实验模型,使用的数据量较少,训练也较少.
由于AI-Toolkit不好用,我后续切换至OneTrainer.
FAQ
Comments (16)
W Furry LoRA. Sad I can't generate this at native 2K resolution though
Too much biased to penis, it's hard to generated female furry😢
Try using other nodes for disabling restrictions or LoRa.
To be honest, the model basically didn't learn the specific content of nsfw at all here. It didn't even learn the main features of my nsfw data, not even the main canine data.
I might just have guided the model to activate its own capabilities. According to my tests, reducing the Lora weight actually performed better in terms of shape quality.
In the nsfw aspect, I don't have much data and added it relatively late. However, during the initial preparation, I considered data balance, and the quantity won't differ too much. But in the overall training data, the female data is indeed insufficient.
@hxyy got it
Having opposite problem lol, making half penis half vagina, really need a gay model that isn't synthetic data slop
@hxyy Same experience
@GPUPoorChad I can't do this well because I only have a 5060 Ti; I'm limited to training at 512 resolution with low model precision.
NSFW content isn't the focus of my training.
@hxyy Sorry didn't mean to come off as asking you to train something, curious though, how do you train things with that little VRAM? I can't train anything with 16GBs, but I think it might be more due to having an AMD card
@GPUPoorChad I can use a 4-bit quantization model for training.
@hxyy AMD is so unoptmized for training even that doesn't seem to help really lol, really need to bite the bullet on a NVIDIA card maybe used 4080
@GPUPoorChad ive thought about doing a NSFW furry finetune since I have the hardware but I am waiting for generalized NSFW finetune to surface so i do not have to collect 100k+ images to have it learn basic NSFW anatomy let alone NSFW+Furry from scratch. In captioning alone for a dataset that size it would be billions of input tokens.
@MisticRain69 Looking forward to that, keep us updated if possible :O
It is specifically biased to canine penises. I can't get it to make a penis without a knot. (this reveals something about the training set, lol)
@Acleveralias Indeed, the tendency towards the new Lora model is too strong because it is difficult to control the actual training of concept learning.
Initially, it would lean towards humanoid during training, then it would be a mixture of semi-human and semi-canine, and later, although it leaned towards canine, even though I used human data, it was impossible to avoid it. Although there were horses in the data, the learning of horses was the most difficult.
After multiple tests with the small Lora model, even if the number of horse images was twice that of canine images, it would eventually develop canine-like features.
However, the original model had a strong human tendency, so the Lora intensity could be reduced.
(Machine translation)
This is an excellent model. I was surprised to find that the model itself can act as an NSFW switch for Krea2, independently enabling NSFW mode without the need for "Krea2FilterBypass" (though I have not tested it with female characters). I am curious whether this is due to specific parameter configurations or simply because the training data includes a small amount of NSFW content？
Just because it contains a small amount of NSFW content, its performance in the NSFW aspect is very unstable. This merely activated the model's own capabilities, and it didn't learn the main canine features from me.
I didn't collect enough data, and the AI-Toolkit is relatively slow in training. This model is just a trial. I will retrain it in OneTrainer.
However, due to the model's own data bias, the lack of specific knowledge, and my inability to prepare too much data to handle different situations, it is difficult to control the learning between the overall or specific details in this regard. This is because what I want to do is to create a model that covers a wide range of styles. If it's just targeted learning of features, it is relatively easier.