(10/01/2024) IMPORTANT INFORMATION UPDATE:

I have uploaded a new model in the files section called dvhairy-000009.zip. It is a safe-tensors file that I would love for people to have a go with and give me some feedback if they are interested (I had to zip as it wouldn't allow two safe-tensors of the same size in the same version). It has been heavily updated, is reasonably functional, but requires significantly more training. Changes are:

• Keyword is no longer required, can go straight into normal prompting as long as the Lora is active, however, it has not been trained on words like 1girl, solo or woman so these may yield normal checkpoint results. It has been trained on phrases such as, anus, pussy, breasts, from behind, spread legs etc...so any of those should bring the Lora into effect. If you wish to use a keyword it is simply dvhairy, but it is really not required.

• It has also been trained on 3 new individual concept words: Ginger, Dildo and PeePee. They are all self explanatory, but due to the current under training, they do not all converge especially well. Ginger and Dildo work OK, ginger and peepee will often work OK, but dildo and peepee will often present as one or the other. Also, the data for ginger was somewhat limited so a lot of the time the individual may look very similar unless other characteristics are expressly prompted, for example, "ginger, dildo, short hair, pussy, anal pubic hair, from behind" is all that is required for the new sample image I have updated today to illustrate the update.

• The current under-training means some details, particularly butt-holes, are not fully detailed, this is in an attempt to avoid previous training issues where the pace of the training would have resulted in over-fitting of the model if I trained more steps, but at the same time, the lack of steps resulted in anatomy mismatch. This was exasperated by an imbalance in training data of different concepts meaning that certain patterns were under represented during the learning process. I have dampened the learning process to allow more steps and also updated the data set to accommodate the new concepts and re-balanced it to help facilitate the learning process.

• I will be updating the data set to include a higher representation of ginger pubes to hopefully rectify the current issue where their is often a deviation from the expected image generation, and to hopefully, match armpit hair. If anyone knows if color augmentation could be causing this this, I would love to hear from them, particularly before updating the lora to include a blonde pubic hair concept.

• I am aware of an issue where the ginger pubic hair is sometimes recognized by the AI as orange/yellow panties. I am hoping further training will resolve this.

• the dildo object is known by the AI and since it was tagged in training it stays in the style that the AI has learned, so all dildos just look like a big pink or purple piece of plastic. This is not all bad, and the lora does work, however, when updating I will be removing the dildo tag to teach the AI the style I would like. As such the dildo keyword will be replaced with a currently unknown concept. I will update details of this in the next version.
  

Again, any feedback is appreciated. I am also more than open to taking requests for concepts people would like included as it seems that every-time the data set is increased, the style outcome is better since there is more for the AI to learn from as long as the overall hairy style is maintained. Saying that, please note that object insertion (bats, bananas, cucumbers and bottles) which I am already training, will instead be included into my first major checkpoint which I am hoping to have enough training done for over the next few months. :)

(Original Post) PLEASE READ:

This is my first attempt at a style. My end goal is to merge a mass of styles into a few mega styles, then eventually a checkpoint. I will be mainly focused on the kinky side, but will attempt to adhere to realism as much as I can. I am particularly focused on trying to achieve significantly detailed and realistic genitalia across the most diverse range of body types, ages and ethnic groups as possible. As a team red player, I am also going to achieve this using only my own graphics card, an AMD Red Devil 16GB 6900XT (Which I plan on upgrading next generation).

To the end of my main goal, I am open to all criticism, comments and suggestions, all I ask is that they are constructive and polite, which I find generally people of this community tend to be anyway.

Anyway, I have tested this across multiple checkpoints, using randomly generated prompts to ensure as accurate testing model diversity as possible. Below are a few things of note (Advice on how to possibly rectify these would be appreciated):

At full strength the lora does have a tendency to produce multiple assholes, particularly when using the "from behind" angle. This seems to occur far less in the RealCartoon-Realistic checkpoint. Dropping to 0.8 strength also seems to help significantly.

When used to generate an image of an especially overweight individual, the size of the labia and the distribution of body fat causes some distortion, particularly in "close-up" or "spread pussy" shots.

My best generations so far have been from the popular real cartoon realistic checkpoint. Lazymix+, PicXReal and CyberRealistic have also produced good results.

Has been trained on A LOT of images. I have manually resized and cropped the images before training to ensure all watermarks have been removed and that the image focus area is correctly set. I have also experimented with Max Norm Regularization in an attempt to ensure compatibility with other loras. This has not been extensively tested in this initial version so any feedback is appreciated.

The activation text is "atkstylehairy" but other keywords have also been trained: "close-up, from behind, anus, pussy, armpit, spread pussy, fingering, spread anus...etc." Experiment.

Whats Next for DangerVine:

atkstylepetite

atkstylepov

These are both pretty self explanatory. I am also in the process of currently upgrading my self anal fisting model to include more styles of self fisting and POV fisting. I am upgrading it to an objectinsertation lora, which will have multiple styles, so far it includes bats, bananas, cucumbers, bottles, dildos and fists. It will be trained to work with my atkstyle loras specifically. The atkstylepov will include penis training for proper POV viewing, this an area I always find lacking and I will be focusing on training a high quality penis that doesn't come at the expense of every generated individual having to have one.