Amber is a LoRA for Pony models that you can use to generate realistic images of beautiful women with pale skin.

Recommended usage

1. Put "score_9, score_8_up, score_7_up, realistic, ccamber, 1girl," at the beginning of the positive prompt.

2. Set LoRA weight to 0.8 and tweak from there.

3. Use highres fix.

4. If freckles are too much (especially when you use ComfyUI), specify the word "freckles" in negative prompt and adjust the weight of this word.

5. Use with CyberRealistic Pony v7 as the main model (https://civarchive.com/models/443821/cyberrealistic-pony) as it was used for LoRA training.

Example prompt

Positive

score_9, score_8_up, score_7_up, realistic, ccamber, 1girl,

portrait,

22yo woman, (beautiful, sweet),

dress, mansion

<lora:ccamber_v2:0.8>

Negative

score_6, score_5, score_4, ugly

How this LoRA was created

Dataset creation

Main dataset

• 115 synthetic safe-for-work images of women were generated using the epiCRealismXL v7-Final Destination model (https://civarchive.com/models/277058?modelVersionId=489217) with the words "22yo woman" included in the prompt.

• Faces were further tweaked using the The Realistic Vision V6.0 B1 (VAE) model (https://civarchive.com/models/4201?modelVersionId=245598) with the words "22yo woman" included in the prompt.

Regularization dataset

• 199 synthetic safe-for-work images of women were generated using the epiCRealismXL v7-Final Destination model (https://civarchive.com/models/277058?modelVersionId=489217) with the words "22yo woman" included in the prompt. Face enhancement was not done for this dataset.

Training

Kohya trainer was run with the following parameters:

accelerate launch --num_cpu_threads_per_process=2 "./sdxl_train_network.py" \
  --enable_bucket \
  --min_bucket_reso=256 \
  --max_bucket_reso=2048 \
  --pretrained_model_name_or_path="<path to model dir>/cyberrealisticPony_v7.safetensors" \
  --dataset_config=<path to dataset dir>/ccamber_v2/ccamber_v2.toml \
  --resolution="1024,1024" \
  --output_dir=<path to LoRA dir>  \
  --output_name=ccamber_v1 \
  --logging_dir="output/log" \
  --network_alpha=128 \
  --save_model_as=safetensors \
  --network_module=networks.lora \
  --text_encoder_lr=0.0001 \
  --unet_lr=0.0001 \
  --network_dim=128 \
  --lr_scheduler_num_cycles="10" \
  --no_half_vae --learning_rate="0.0001" \
  --lr_scheduler="cosine" \
  --train_batch_size="1" \
  --max_train_steps="5000" \
  --save_every_n_epochs="1" \
  --mixed_precision="bf16" \
  --save_precision="bf16" \
  --cache_latents \
  --cache_latents_to_disk \
  --optimizer_type="Adafactor" \
  --gradient_checkpointing \
  --optimizer_args scale_parameter=False relative_step=False warmup_init=False \
  --max_data_loader_n_workers="0" \
  --bucket_reso_steps=64 \
  --xformers \
  --bucket_no_upscale \
  --vae <path to VAE>