lora-dataset-prep

lora-dataset-prep

Turn a folder of raw photos into a clean, captioned, validated LoRA training dataset — in one command.

Quick start · How it works · Captioning · Validation · Agent skill · Reference

Training a LoRA is easy. Preparing the dataset is where it goes wrong: duplicate frames overfit a pose, an uncaptioned uniform becomes part of the character, one repeated adjective makes every generation warm-lit. This tool catches all of that before you spend GPU hours.

Built for Ostris AI Toolkit / Z-Image Turbo. Character LoRA first; style, object, clothing, and environment supported.

 raw_photos/                      my_dataset/
 ├── IMG_4032.jpg                 ├── train/
 ├── фото (3).png        ──►      │   ├── 01.png • 01.txt   caption sidecars
 ├── screenshot.webp              │   ├── 02.png • 02.txt
 └── ...                          │   └── ...
                                  ├── raw/                  originals, untouched
                                  ├── README.md             auto-generated dataset card
                                  └── metadata.json         full provenance

Quick start

No dependencies needed for the core workflow — Python standard library only.

# 1 · Dry-run: prints the full plan, changes nothing
python prepare_dataset.py --source ./raw_photos --output ./my_dataset \
    --type character --trigger mychar01

# 2 · Apply
python prepare_dataset.py --source ./raw_photos --output ./my_dataset \
    --type character --trigger mychar01 --execute

# 3 · Hand-edit captions, then re-check (read-only)
python prepare_dataset.py --validate-only --output ./my_dataset

Prefer questions over flags? Just run python prepare_dataset.py.

How it works

Every run is a dry-run by default. Files change only with --execute.

Caption modes

vlm needs pip install anthropic + ANTHROPIC_API_KEY. Raw BLIP output violates the caption formula and must be rewritten — details in reference.md.

The captioning rule

Describe what you do not want the LoRA to learn. Everything left out of the caption gets bound to the trigger token.

So the face is never described — and clothing, pose, lighting, and background always are, keeping them flexible at inference:

+ zvqmark, a still from a TV show, medium shot of a man wearing a dark police
+ uniform with a blue shirt and tie, standing in a kitchen with hands on hips,
+ looking down with a stern expression, warm interior lighting

- zvqmark, masterpiece, best quality, handsome man, detailed face, 8k

The full guide — formula, include/omit tables, per-type strategies, typical mistakes — lives in reference.md.

What gets validated

Structure — image↔caption pairing, orphans, duplicate basenames, unsafe filenames

Captions — trigger presence and position, length (15–35 word target), unfilled [BRACKET] placeholders, filler phrases, multiple subjects, failed-VLM markers

Attribute sticking — any content phrase repeated in ≥ 50 % of captions is flagged; the silent killer that bakes one lighting or gaze into every generation

Images — minimum dimensions, blur heuristic (tunable via --blur-threshold, 0 disables), EXIF rotation flags (phone photos that would train sideways — --resize bakes the rotation into the pixels), exact and near-duplicate detection

Use as an Agent Skill

Drop the folder into your skills directory — the agent picks it up whenever you ask it to prepare a dataset, write captions, or choose a trigger word:

The agent reads SKILL.md for the workflow and reference.md for captioning rules, then drives the CLI for you.

Options

Complete table: SKILL.md → Options Reference.

Development

pip install pytest Pillow
python -m pytest test_prepare_dataset.py -q        # 79 tests, ~3 s

The suite covers caption generation, hashing, perceptual dedupe, resize, EXIF-orientation handling, every validation rule, attribute-sticking detection, and the VLM path (mocked API — no key needed). CI runs the same suite on Python 3.9–3.12 for every push; releases are tagged and summarised in CHANGELOG.md.

lora-dataset-prep/
├── SKILL.md                   agent entry point + full options table
├── reference.md               captioning & dataset guide
├── prepare_dataset.py         the tool — stdlib-only core
├── test_prepare_dataset.py    test suite
└── LICENSE                    MIT

MIT © afloy011-spec

If this saved you a failed training run, a star is appreciated.