This is a merge of two partially-trained snapshots from my Fluxxxible and Reclining Nude, After Sex LoRAs which I have been using to get a head-start for training LoRAs with more complex NSFW concepts. I am sharing it to help accelerate your training process and improve the detail of female nipples and genitals.

This LoRA is not meant to be used on its own and will produce bad hands, strange candy nipples, and other glitches. That is intentional, allowing you to specialize it with further training.

If you train any LoRAs using this base, please let me know and I will add some of them to the suggested resources.

Training

In my testing, training more complex NSFW concepts on top of this LoRA can be accomplished within 500 steps or so. I prefer using medium-sized datasets of about 250 images, but I have had successful results with micro-sized datasets of 20-25 images as well.

This LoRA was extracted at rank 32 and should be trained at the same rank for best results. I thoroughly tested ranks from 8 to 128, and this provided the best results without a large loss of quality.

I prefer to train using an image size of 1024x1280, with images converted before training to allow better conversion algorithms. That allows a batch size of 6 images when used with an A6000 or A40, and training takes < 6 hours for most configurations, or about $2-3 when using RunPod.

My hyperparameter settings are usually like this:

d_coef: 2
learning_rate: 1.0
learning_scheduler: cosine
max_epochs: 15
network_alpha: 32
network_dim: 32
noise_offset: 0.1
num_cycles: 1
optimizer_type: prodigy
optimizer_args: |   
  "decouple=True" 
  "weight_decay=0.1" 
  "betas=0.9,0.999" 
  "use_bias_correction=False" 
  "safeguard_warmup=False" 
resume_from: extracted-fluxxible-reclining-60-40.safetensors 
save_every_n_steps: 50 
sample_every_n_steps: 100 
timestep_sampling: shift 
weight_decay: 0.1 

My dataset settings are usually like this:

{
        "datasets": [{
                "resolution": 1280,
                "enable_bucket": true,
                "min_bucket_reso": 768,
                "max_bucket_reso": 1280,
                "bucket_reso_steps": 256,
                "bucket_no_upscale": true,
                "batch_size": 6,
                "subsets": [...]
        }]
}

More details to come soon. I will be posting more LoRAs using this base over the next few weeks and I am hoping to write an article with more training configurations when I have some time.

If you have any suggestions on improving this process, please let me know. I have tested a large variety of hyperparameter combinations over multiple weeks of training time, and this is the best combination that I have found so far, but I am happy to make it better.

Prompts

The prompt format for this LoRA was based on the Llama/Joy captioning style, with some words being censored, as seen in my other LoRAs. You will have the best results if you use a similar prompt format with good detail, but I have trained some LoRAs using very short prompts, even single keywords.

The keywords that have been censored include:

• 4nus for anus

• 4n4l for anal

• 4ss for butt, buttocks

• br34st for breast and breasts

• cl1t0r4l for clitoral hood, etc

• cl1t0r1s for clitoris

• d1ld0 for dildo and sex toy

• n4k3d for naked

• n1ppl3 for nipple and nipples

• nud3 for nude

• 0r4l for oral

• 0r4l s3x for oral sex

• p3n1s for penis

• s3x for sex

• v4g1n4 for vagina