Character LoRA Creation Guide & Live Workflows

Author: floyoofficial

Table of Contents

Step 1: Create/Gather Dataset Images

The foundation of any high-quality LoRA is a diverse dataset. Your training images should include:

Varying angles - front, side, back, three-quarter views
Different expressions - happy, serious, surprised, etc.
Multiple poses - standing, sitting, action poses
Various lighting conditions - bright, dim, natural, artificial
Different environments - indoor, outdoor, various backgrounds

Why Variety Matters: Diversity in your dataset is crucial because LoRAs learn patterns from your training data. If your dataset contains too many similar images, the LoRA may "bake in" repetitive characteristics that become difficult to override during generation. For example:

Too many images with the same pose → LoRA always generates that pose
Repetitive facial expressions → Limited emotional range in outputs
Consistent lighting/backgrounds → LoRA struggles with different environments
Similar angles → Poor performance when requesting different viewpoints

A varied dataset ensures your LoRA remains flexible and responsive to different prompts rather than being locked into learned patterns.

Dataset Requirements:

30-100 images at 1024x1024 resolution (optimal for Flux)
More images can improve results, provided they maintain variety
Place all images in a single folder

Alternative: Synthetic Dataset Generation If you have a single character design, you can use the Image to Character LoRA Dataset workflow with Flux Kontext to generate a diverse training set automatically.

For this process, you may bring your own character design or create a new one using the Text to Image workflow.

⚠️ Important: When using AI-generated images, manually review and remove any images with:

Deformed anatomy
Nonsensical backgrounds
Visual artifacts or errors

These flaws can be learned and reproduced by your LoRA if left in the training data.

Step 2: Caption Your Dataset

Captions are text descriptions that accompany each training image, teaching the LoRA what it's looking at. These descriptions help the model understand which visual elements correspond to your character versus background elements, clothing, poses, and other scene details. Quality captions are essential because they guide how the LoRA learns to respond to prompts - poorly described images can lead to confused outputs or unwanted associations being learned alongside your character.

Captioning can be done manually by creating text files and writing descriptions of each image, making sure that the file names match (image_01.png, image_01.txt), repeat for each image in your dataset.

Alternatively, you can generate captions automatically with the Detailed Auto Caption workflow:

Input your dataset folder path
Set your desired output path
The workflow will automatically:
- Resize images to 1024x1024 (or appropriate aspect ratio)
- Generate detailed captions for each image
- Save both processed images and matching caption files

This prepares your dataset for training with minimal manual work required.

Step 3: Training Your LoRA

Time to train! Use the ComfyUI Flux LoRA Trainer workflow:

Setup Requirements:

Input your captioned dataset folder path (in all three workflow inputs)
Set your trigger word in all designated nodes
Create a sample prompt for testing
Name your LoRA

Training Process:

Click generate and wait 1.5-3 hours (depending on dataset size)
The training will produce 4 checkpoint benchmarks:
- 750 steps
- 1500 steps
- 2250 steps
- 3000 steps (final)

Step 4: Test and Compare Results

Congratulations! Your LoRA is trained. Now determine which checkpoint works best using the Flux Character LoRA Test and Compare workflow:

Load all 4 LoRA checkpoints (750, 1500, 2250, 3000 steps)
Apply your trigger word
Use the provided example prompts or create your own (trigger word must be included)
Generate a comparison grid of 16 images

Evaluation Tips: