So I just started messing with Dreambooth using Lora, and HOT damn I have probably the best model ever.. but still in development. Long story short, I was tired of seeing ugly belly buttons, and always hands-on or around the belly. That plus textual inversion and it's like holy crap man. Most of these were done with 1 pass, I might of touched up some because I am building a "class folder" that Dream booth will use to compare with when learning new models.

Here are some photos as an example, still a work in progress:
A beautiful pregnant woman in a nude at the beach, preggo,  sexy,  high quality photography, 4k, masterpiece, color, Canon EOS R3, F/22
Negative prompt: animation, anime, cartoon, extra fingers, extra limbs, ugly, mutation, deformed, fused fingers, man, child, old, wrinkles, animation, anime, cartoon, extra fingers, extra limbs, ugly, mutation, deformed, fused fingers
Steps: 20, Sampler: DPM++ 2S a, CFG scale: 7, Seed: 1078548715, Face restoration: CodeFormer, Size: 768x768, Batch size: 2, Batch pos: 1, Denoising strength: 0.7, First pass size: 0x0

Nice, we need a good realistic pregnancy model for sure. I haven't tried LORA yet, but I like the results
What is the dataset size and learning rate?

How big is the HyperPregnancy V2 model?

Would it be possible if someone could share it via mega? I don't really like using torrents.

(December 2, 2022, 7:17 pm)thojm Im in the process of uploading them all to https://civitai.com/ so give me just a bit of time

Edit: alright all of my hyper models have been uploaded https://civitai.com/?username=throwawayjm  You may have to make an account to download, but it's an alternative option to torrenting at least

The model is hosted here now.  But the site requires an account to download NSFW models.

Is that model only for anime though? Also thanks for that site link. I've been trying to find more nsfw models.

v1.1 is the realistic, one but it's not great at it since it was trained with anime style images. Hard to find enough real hyperpreg

I'm hoping to see your Lora version of Hyperpreg. That way we can have different sizes, etc. Which by the way how do you make a Lora version?

Been trying to train one for a while, but only just got a good version going. Ill upload the model soonish.  Eventually might try training a realistic vanilla pregnant LoRA when I get some time.
I train LoRA models with this https://github.com/kohya-ss/sd-scripts

Edit: I should mention LoRA is way quicker and easier to train than Dreambooth so far, its just this one particular model that was giving me trouble.

Do you have a good step-by-step guide for training on larger training sets? I used aientrepeneur guide for Kohya, but my stuff was over-trained. I have a larger dataset of over 200 images, but he explicitly said to do 100 steps per image. Any help would be great. Maybe you can post your Json settings? I think you used Dreambooth too, so if you can specify those settings too, I also would appreciate it!

Been training with Kohya's LoRA trainer lately.  But the config below should be similar enough for Dreambooth.
The trick to prevent overtraining is to save chekpoints often, like every few epocs, and just test earlier versions if the last one is overtrained

Quote:#for 100-200 images

export short_name="<image folder name>"
export pretrained_model_name_or_path="../models/anything-v3"
export vae="../models/Anything-V3.0.vae.pt"
export raw_img_folder="${short_name}"
export steps=1000
export learning_rate=8e-5  #1e-4 is default
export train_batch_size=2
export gradient_accumulation_steps=8
export mixed_precision="bf16" # leave fp16 if not on Ampere GPU
export num_cpu_threads_per_process=6
export max_resolution=768
export max_resolution_bucket=768
export LR_SCHEDULER="constant"
export network_dim=32 # many people use 128, but I like to try and squeeze all the knowledge into smaller files
export WARMUP_STEPS=$((${steps} * 10/100)) # 10% total steps

FULL_NAME="${short_name}_<other params you want to append to output name>"

accelerate launch --num_cpu_threads_per_process $num_cpu_threads_per_process train_network.py \
    --pretrained_model_name_or_path="${pretrained_model_name_or_path}" \
    --train_data_dir="../images/${raw_img_folder}" \
    --output_dir="./output/${short_name}" \
    --resolution=$max_resolution \
    --train_batch_size=$train_batch_size \
    --gradient_accumulation_steps=$gradient_accumulation_steps \
    --gradient_checkpointing \
    --learning_rate=${learning_rate} \
    --lr_scheduler="${LR_SCHEDULER}" \
    --max_train_steps=$steps \
    --use_8bit_adam \
    --xformers \
    --mixed_precision=$mixed_precision \
    --save_every_n_epochs=1 \
    --save_model_as="safetensors" \
    --clip_skip=2 \
    --seed=42 \
    --flip_aug \
    --network_module=networks.lora \
    --vae="${vae}" \
    --enable_bucket \
    --max_bucket_reso=$max_resolution_bucket \
    --shuffle_caption \
    --logging_dir=logs \
    --log_prefix="${short_name}_lr${learning_rate}_" \
    --output_name=$FULL_NAME \
    --network_dim=$network_dim \
    --network_train_unet_only