Why LoRA Training Matters for Professional AI Content
Training your own LoRA (Low-Rank Adaptation) model is essential when you need a consistent character or style across your AI-generated content. Without a LoRA, each generation produces a different face, body, or aesthetic — making it impossible to create professional content like music videos, social media campaigns, or branded materials.
This becomes especially critical when you use upscaling with detail enhancement workflows. Standard upscaling can drift your character’s appearance. With your trained LoRA, even the fine details — skin texture, eye color, facial structure — remain consistent throughout all outputs.
Once trained, you can use your LoRA with any open-source workflow on Kitty AI Studio — from text-to-image generation to video creation and beyond.
This comprehensive guide will walk you through the entire LoRA training process — from creating your dataset to testing your finished model on druidcat.com.
What You’ll Learn
- How to build a complete dataset in one click with the 1-Click Dataset Creator
- How to do it manually with Nano Banana 2 or GPT Image 2 if you prefer full control
- Generating captions automatically (or with Florence / ChatGPT)
- Setting up Runpod with the Ostris AI Toolkit template — the trainer we use
- Training LoRAs for Z-Image and WAN 2.2 architectures
- Testing your completed LoRA on druidcat.com
Training Plan Overview
Here’s what we’ll cover in this guide:
- Create a Diverse Dataset — one click with the Dataset Creator, or manually with Nano Banana 2 / GPT Image 2
- Generate Captions — automatic in the Dataset Creator, or Florence / ChatGPT by hand
- Launch Runpod Training — using the Ostris AI Toolkit template
- Test Your Results — upload and use your LoRA on druidcat.com
Step 1: Creating Your Dataset
A LoRA is only as good as its dataset. You need the same person seen from many angles, with many expressions, in different places and outfits — otherwise the model learns one photo instead of one character.
There are two ways to do this on druidcat.com. Start with the fast one.
The Fast Way: 1-Click Dataset Creator
Upload one photo of your character, pick how many images you want, and the 1-Click Dataset Creator generates the whole training set for you — close-ups, profile views, full body poses, expressions and location shots — using either Nano Banana 2 or GPT Image 2.
Every prompt is pre-written and fully editable: rewrite any line, or remove the ones you don’t want. Each prompt carries a built-in identity lock, so the model changes the pose and the scene but keeps the same face.
It also does the boring part: enter your trigger word and you get a ready-to-train ZIP — images named Olivia (1).png, Olivia (2).png… each with a matching .txt caption file. That is exactly the folder structure the trainer expects, so you upload it to Runpod and start training. No renaming, no captioning scripts, no Python.
Pricing is simple: you pay per generated image (the normal price of the model you picked), plus a small flat fee per caption if you enable automatic captioning. The panel shows the exact total before you start.
The Manual Way: Build It Yourself
If you want to art-direct every frame, you can build the dataset by hand. The workflow below is the same one the Dataset Creator automates.
1. Generate the face
Go to Kitty AI Studio and start with a close-up portrait — close-ups carry the most facial detail, which is what the LoRA needs most. Good starting models:
- Z-Image Turbo — fast, great aesthetic, open source
- Krea 2 Turbo — the most realistic open-source option right now
- Instagirl — influencer-style aesthetics
- Qwen Text to Image — natural beauty
Write a detailed prompt. Say “close up” for a face shot, then describe the features. For long hair, a medium shot often works better. Any chat model can help you sharpen the prompt.
close up portrait, young woman with emerald green eyes, auburn wavy hair, soft natural lighting, professional photography, detailed skin texture, slight smile
2. Expand into angles, poses and scenes
Take that portrait and feed it to an image-edit model that keeps identity. Two work well, and both are one-step — you no longer need a chain of separate angle / inpaint / clothes-swap workflows:
- Nano Banana 2 — Google’s editor. Excellent identity retention, understands camera direction, and changes outfits, poses and locations from a plain sentence.
- GPT Image 2 Edit — very strong prompt following, good for precise expression and framing instructions.
Cover these categories — this is the shot list that actually matters:
- Face, frontal — extreme close-up, neutral expression, looking into the camera
- Face, angles — 3/4 view, full 90° side profile (left and right), slight high angle, slight low angle
- Expressions — gentle smile, open-mouth laugh, sad, angry, surprised, crying, talking mid-sentence, soft “pretty eyes” look
- Full body — frontal A-pose on a neutral background, back view, side view walking
- Medium shots — seated, leaning, hands visible
- Locations — café, desert, city street at night, beach, forest, office, hotel room, studio
- Outfits — a few clearly different clothing styles
Use at least 2000 px for full body shots so the face survives; close-ups are fine at 1K and cost less.
Dataset Size Recommendation
20–40 images is the sweet spot for a character LoRA. Below 15 the model struggles to generalise; above ~50 you mostly pay more for very little gain.
A good split: ~40% face close-ups (this is what makes the likeness), ~35% medium shots, ~25% full body. Vary background and outfit as much as you can — if every photo has the same white wall, the LoRA learns the wall too.
Step 2: Creating Captions
Captions are .txt files sitting next to each image, describing what’s in it. They are essential for training: they teach the model which words map to what it sees, so your LoRA responds correctly to prompts later.
- One
.txtper image, with exactly the same filename —Olivia (1).png→Olivia (1).txt - Start each caption with your trigger word, then a comma — e.g.
olivia1, a portrait of a young woman with long red hair… - Describe what varies (pose, outfit, location, expression, framing). Don’t re-describe what never changes about the person — that’s what the trigger word is for.
- Plain descriptive sentences beat keyword soup.
Option A: Automatic (recommended)
The 1-Click Dataset Creator captions the whole set for you and puts the .txt files in the ZIP, correctly named and optionally prefixed with your trigger word. It costs a few cents per caption and saves the entire step.
Option B: Florence Script (Windows, free)
A local Florence script — no cloud, no cost:
- Install Python from python.org. During installation, check “Add Python to PATH” — this is critical!
- Unpack the Florence folder anywhere on your computer
- Open a terminal in that folder: click the address bar in File Explorer, type
cmd, press Enter - Install dependencies:
pip install -r requirements.txt - Run it:
python app.py, then open the link shown in the terminal
In the interface: drop your images in, put your trigger word in Caption Prefix (e.g. “olivia1”), choose Base (faster) or Large (more accurate), and download the ZIP of caption files.
Option C: ChatGPT or Gemini
Upload the images and ask for one caption file per image, same filenames, trigger word first.
Create caption text files for these images. Each file must have the same name as the image but with a .txt extension. Start every caption with "olivia1," followed by a detailed description of pose, outfit, expression, framing and location.
Step 3: Training on Runpod
Training needs a real GPU for a few hours, so we rent one. We use Runpod with the Ostris AI Toolkit template — that template is the trainer itself (a web UI for ai-toolkit), which is why the steps below tell you to search for it explicitly.
Create your account first: Sign up for Runpod →
You pay per hour only while the pod runs. A character LoRA typically costs a few dollars of GPU time in total.
3.1 — Deploy the pod
| Step | What to do |
|---|---|
| 1 | In the Runpod dashboard, open Pods |
| 2 | Switch to Community Cloud (cheaper than Secure Cloud, fine for training) |
| 3 | Pick a strong GPU — RTX 5090 or RTX 6000 recommended |
| 4 | Click Change Template and search for “ostris” — this is the important one |
| 5 | Select the Ostris AI Toolkit template (listed as AI Tools) |
| 6 | Training more than one model? Click Edit Template and raise the container volume to 250 GB |
| 7 | Click Deploy and wait for the pod to finish loading |
3.2 — Connect and upload your dataset
| Step | What to do |
|---|---|
| 1 | Click Connect on your running pod |
| 2 | Open port 8675 (the AI Toolkit web UI) |
| 3 | Password: password — the template’s default |
| 4 | Go to Datasets and upload your whole folder — images and .txt captions together |
.txt means that image trains with no caption at all.
3.3 — Configure Z-Image training
Open the New Job tab and set:
| Setting | Value | Notes |
|---|---|---|
| Training Name | influencer_name_zimage | Your LoRA filename |
| Trigger Word | olivia1 | Must match your caption prefix |
| Model Architecture | Z-Image De Turbo | For Z-Image LoRAs |
| Linear Rank | 32–44 | 44 for a stronger LoRA |
| Training Steps | 2500–3000 | Start with 2500 |
| Resolutions | 1024, 1280 | Enable both |
In the Sample tab:
- Remove all example prompt boxes except one
- Paste a prompt from your captions, but move your character into a new scenario (e.g. “sitting in a café”)
- These sample images are how you spot overtraining early
Click Create Job, then hit the Play button (triangle) in the Dashboard.
3.4 — Settings for WAN 2.2 (14B)
The differences versus Z-Image:
- You get TWO files: High Noise and Low Noise — download both!
- In the Sample tab: set Number of Frames = 1 and FPS = 1
- Enable Skip First Sample (we’re training for images, not video)
- Everything else stays the same
After a few hours your LoRA files appear in the Dashboard under Checkpoints.
Step 4: Testing Your LoRA
Head back to Kitty AI Studio and try it out.
Testing a Z-Image LoRA
- Open Z-Image Turbo
- Write your prompt including the trigger word
- Tick the custom LoRA box and upload your file
- Generate
Testing a WAN LoRA
- Open WAN 2.2 Text to Image
- Add BOTH the High Noise and Low Noise files
- Include your trigger word in the prompt
- Generate
Where Can You Use Your LoRA?
Your trained LoRA works across many workflows on Kitty AI Studio:
- Instagirl — Instagram-style photos
- Z-Image Detailed — high-detail images with upscaling
- Image to Video — bring your character to life
- SVI Pro Extended — long-form video with your character
- And many more!
Conclusion
You now have a complete toolkit for creating AI influencers. Build the dataset in one click, caption it automatically, train on Runpod with the Ostris AI Toolkit, and use the result anywhere on Kitty AI Studio. The key is consistency — a well-trained LoRA keeps your character identical across every generation, which is exactly what makes the content sellable.
Ready to start? Open the 1-Click Dataset Creator →
