Flux LoRA Dataset Preparation Tool

Prepare image datasets for Flux LoRA training with easy AI captioning, PNG conversion, and flexible output options.

---

Features

• Convert JPG/JPEG images to PNG format
• AI-powered captioning (using BLIP or filename)
• Save captions in .txt and/or .json formats
• Simple CLI (prepare_flux_lora_dataset.py)
• Modern GUI (flux_lora_dataset_gui.pyw) — with Pause, Resume, and Stop!

---

Installation

Install the required dependencies:

pip install -r requirements.txt

---

Usage

1. Command-Line (CLI)

Process all images in a directory using BLIP captioning:

python prepare_flux_lora_dataset.py input_directory output_directory

Options

• --tagger {blip,simple}: Choose caption generator
  • blip: Uses BLIP (AI; recommended)
  • simple: Uses filenames (fast, no AI)
• --format {txt,json,both}: Caption output format
  • txt: .txt files (default for most trainers)
  • json: .json files
  • both: Both
• --no-convert: Skip PNG conversion (keep originals)
• --device {auto,cuda,cpu}: Device for BLIP
  • auto (default): Use GPU if available, else CPU
  • cuda: Force GPU
  • cpu: Force CPU

Examples

Basic BLIP (both formats):

python prepare_flux_lora_dataset.py ./images ./processed --tagger blip --format both

TXT only:

python prepare_flux_lora_dataset.py ./images ./processed --format txt

No AI (filename tags):

python prepare_flux_lora_dataset.py ./images ./processed --tagger simple

Don't convert PNGs:

python prepare_flux_lora_dataset.py ./images ./processed --no-convert

---

2. Graphical User Interface (GUI)

The GUI makes dataset preparation easy and interactive. You can pause, resume, or stop processing — ideal for big folders!

Start the GUI

python flux_lora_dataset_gui.pyw

Key Features

• Folder selection with file browser
• Choose tagger and output format
• Optional PNG conversion
• Live log and visual progress bar
• Pause/Resume/Stop controls
• Error reporting and summaries when done

---

Output Format

TXT Files

Each image gets a .txt file with the caption:

image001.png
image001.txt  (contains: "a beautiful landscape with mountains and trees")

JSON Files

Each image gets a .json file with:

{
  "image": "image001.png",
  "caption": "a beautiful landscape with mountains and trees",
  "source": "image001.jpg"
}

---

Notes & Tips

• Supports .jpg, .jpeg, and .png files.
• BLIP model downloads automatically (~990MB, first run).
• Use GPU (cuda) for much faster BLIP processing.
• Large datasets may take time; use the GUI for easier management.

---

Troubleshooting

Out of memory?

• Use --device cpu or select CPU in the GUI.
• Split your dataset into smaller batches.

BLIP doesn't work?

• Ensure you have transformers and torch:
  pip install transformers torch
• Check for enough disk space for the model.

Image conversion errors?

• Make sure your images are valid (not corrupted).
• Check permissions for the input/output folders.

---

Both CLI and GUI work cross-platform. For best results, use the GUI for large or complex batches!