Generation Guide

Model Information

• Model Name: {model_name} (replace with the actual filename you downloaded, e.g., gngsfimZIB.safetensors)

• Trigger Word: {trigger_word}

Recommended Settings

Resolution

• 2:3 ratio: 821×1232 (portrait)

• 3:2 ratio: 1232×821 (landscape)

• Square: 1:1

• Note: You can vary these resolutions with limited success

• FT15 models: Lower max resolution at 512×768

Generation Parameters

• Sampler: Euler (typically)

• CFG Scale:

  • Standard models: 3-7

  • Turbo models: 1

• Steps:

  • Standard models: 20-50

  • Turbo models: 9

• LoRA Strength: 0.6-1.0

  • If images look "cooked" or overprocessed, lower the strength

Model Series Identifiers

• FT15 - Stable Diffusion 1.5 (max resolution: 512×768)

• XLrd - SDXL Run Diffusion X based

• CHHD - Chroma models

• ZIMG - Z-Image Turbo

• ZIB - Z-Image Base

• FKFB - Flux Klein 4B

• QWN - Qwen

Note: LoRA files are large and can be resized if needed

Current Recommendation (January 2026): Use ZIB/ZIT or Chroma models for best results.

Dataset Type Indicators

• mx - Vastly larger datasets with less consistency, typically trained at lower learning rates for longer durations

• lncc - Smaller, more specific aesthetic-focused datasets

Training Data Scale: Datasets vary from 20-30 images to over 1,000,000 images. The median dataset size is closer to 10,000 images.

Training Techniques: Models starting at SDXL use mixed resolution training, multi-subject crop, and flips for improved generalization.

Using the Wildcard Prompt Template

The piped string format below is designed for ImpactPack Wildcard Processor or Automatic1111 Dynamic Prompts. Copy and paste it into either extension to generate a new randomized prompt each time, built on the distribution of the training dataset.

Prompt Format

<lora:{model_name}:{0.6|0.7|0.8|0.9|1}> {trigger_word}, {wildcard_tags}

Example:

<lora:gngsfimZIB:{0.6|0.7|0.8|0.9|1}> example_triggerword, {additional|tags|here}

Understanding the Wildcard Tags

• More pipes (|) in a tag group = rarer tags in the training data

• Fewer pipes or repeated options = more common tags with better model performance

• More examples in the training data mean the model is better at that particular task or concept

Manual Usage (without wildcards)

If you're not using dynamic prompts:

1. Load the LoRA manually in your interface

2. Start with the trigger word {trigger_word} at the beginning of your prompt

3. Add additional tags after the trigger word to vary the composition

4. Tags that appear more frequently in the wildcard examples will produce more consistent results

Tips

• Always start with the trigger word (the first tag) for best results

• Check sample images for embedded generation parameters

• Add additional tags to vary composition and style

• Experiment with LoRA strength if results don't match expectations

• Tags with more training examples will be more reliable and consistent

• Reference the sample images on this page for working parameter combinations

FAQ: Dataset Filename & Trigger Word Conventions

What problem does this filename format solve?

The filename is designed to avoid collisions with generic or common names while also serving as a programmatic signal. It encodes both the trigger word and the dataset type, making it easy for scripts and training pipelines to identify and handle the dataset correctly.

Why not use a generic filename?

Generic filenames tend to overlap across projects and environments. This format ensures:

• Uniqueness across datasets

• Clear intent when parsed programmatically

• No ambiguity about dataset content or usage

What do the suffix codes mean?

The suffix in the filename specifies:

• The resolution of the dataset

• The model architecture tier it is intended for

This makes it immediately clear what kind of model configuration the dataset targets and helps avoid compatibility issues.

What does "mx" stand for?

mx means mix. It indicates that the dataset is diverse and vastly larger (potentially hundreds of thousands to over a million images), though less consistent than focused datasets. These models are typically trained at lower learning rates for longer durations to accommodate the dataset diversity.

What does "lncc" stand for?

lncc indicates smaller, more specific datasets focused on a particular aesthetic. These are more consistent but cover a narrower range of content.

How are trigger words determined?

Trigger words are embedded in the dataset and filename structure. They function as activation tokens that help the model recognize and generate content consistent with the training data. Always use the specified trigger word at the start of your prompt for best results.

How large are the training datasets?

Training datasets vary significantly:

• Minimum: 20-30 images

• Maximum: Over 1,000,000 images

• Median: Approximately 10,000 images

Larger datasets (mx) enable broader capabilities but may be less consistent. Smaller datasets (lncc) are more focused and aesthetically coherent.

For best results, always check the sample images on this model page—generation parameters are embedded in the metadata.

v1.0

768*512

512*768

greatly improved, flexible, high quality,

few known issues in odd poses.

<lora:pmbkkFT15:{1|1.2|1.5}> {pmbkk, }{realistic, }{1girl, }{open mouth, |}{solo, |}{cum in mouth, |}{looking at viewer, ||}{cum, }{portrait, |||||}{tongue, ||||||||||}{close-up, ||||||||||}{facial, ||}{cum on tongue, ||}{eyelashes, ||||||||||||}{nose, ||||||||||||||}{after fellatio, |||||}{upper teeth only, |||||||||||||||}{solo focus, |||||||||||||||}{looking up, ||||||||||||||||}{hetero, |||||||||||||||||}{1boy, ||||||||||||||||||}{penis, ||||||||||||||||||}{uvula, |||||||||||||||||||}{bukkake, ||||}{erection, |||||||||||||||||||}{oral invitation, |||||||||||}{half-closed eyes, ||||||||||||||||||||}{saliva, ||||||||||||||||||||}