Lora dataset prep

LoRA Dataset Prep

Prepare a fal/FLUX-compatible portrait LoRA dataset from a raw photo folder.

Defaults

Setting	Default
Size	1024×1024 JPEG
Trigger phrase	ohwx man (ask user; keep unique)
Caption identity token	[trigger]
Filename prefix	subject
Image policy	Keep user-curated images unless visual QA finds a real defect
ZIP layout	flat: name.jpg + matching name.txt

Workflow

Copy and track:

Progress:
- [ ] 1. Setup env + YuNet model
- [ ] 2. Confirm trigger word + source/out paths
- [ ] 3. Crop every source + generate numbered contact sheets
- [ ] 4. Visually inspect every crop; manually correct failures
- [ ] 5. Descriptive captions with trigger
- [ ] 6. Validate pairs + build flat ZIP; report counts

Resolve skill path (project copy preferred):

SKILL_DIR=".cursor/skills/lora-dataset-prep"
[ -d "$SKILL_DIR" ] || SKILL_DIR="$HOME/.cursor/skills/lora-dataset-prep"
[ -d "$SKILL_DIR" ] || SKILL_DIR="skills/lora-dataset-prep"

1. Setup

From the project working directory:

bash "$SKILL_DIR/scripts/setup_env.sh"

Deps: Python 3, opencv-python-headless, pillow, numpy, fal-client, python-dotenv.
Auth for captioning: FAL_KEY or FAL_API_KEY in .env.

2. Confirm inputs

Ask if missing:

• Source image folder
• Output folder (default: ./prepared_dataset)
• Trigger token (default: ohwx man)
• Subject filename prefix (default: subject)

3. Crop every source

.venv/bin/python "$SKILL_DIR/scripts/prepare_all_portraits.py" \
  --src "/path/to/raw_photos" \
  --out "./prepared_dataset" \
  --model "./models/face_detection_yunet_2023mar.onnx" \
  --prefix "subject"

Produces:

• prepared_dataset/images/*.jpg + [trigger] placeholder .txt
• prepared_dataset/contact_sheets/*.jpg
• prepared_dataset/manifest.json

The detector proposes crops; it does not decide whether an image is usable. Do not discard an image solely because detection fails or its face is distant.

Crops target roughly 4.2× the detected face, but never shrink below 1024 native pixels, so a distant face stays as tight as the source allows without being upscaled past 1:1. A face that still looks small after cropping means the source was low resolution, not that the crop misfired.

4. Visual QA and correction (required)

First generate advisory flags:

.venv/bin/python "$SKILL_DIR/scripts/qc_flags.py" \
  --images "./prepared_dataset/images" \
  --model "./models/face_detection_yunet_2023mar.onnx" \
  --pattern "subject_*.jpg"

Writes prepared_dataset/qc_flags.json, including near_duplicate_of flags for burst shots. Treat these flags as review prompts, not automatic rejection decisions.

Then use image vision to open every file under prepared_dataset/contact_sheets/, not only the automated flags:

1. Confirm the intended subject appears correctly in every numbered crop.
2. Open suspicious outputs and their source files at full resolution.
3. Manually recrop wrong-subject, torso-only, clipped, or off-center outputs.
4. Regenerate contact sheets and inspect the corrected cells.
5. Record manual crop boxes in manifest.json.

Apply each visually chosen square crop reproducibly:

.venv/bin/python "$SKILL_DIR/scripts/manual_recrop.py" \
  --source "/path/to/raw_photos/source.jpg" \
  --dataset-root "./prepared_dataset" \
  --output-name "subject_068.jpg" \
  --box 882 320 2422 1860

The helper replaces the crop, records manual_visual_qa in the manifest, and regenerates all contact sheets. Reopen the affected sheet and verify it.

Never use --delete-flagged without visually confirming each deletion. See references/qc-checklist.md.

5. Caption

Prefer descriptive per-image captions over a single shared caption.

.venv/bin/python "$SKILL_DIR/scripts/recaption_dataset.py" \
  --images "./prepared_dataset/images" \
  --zip "./prepared_dataset/dataset.zip" \
  --trigger "ohwx man" \
  --pattern "subject_*.jpg" \
  --caption-mode placeholder \
  --gender male \
  --workers 8

Set --gender to match the subject. Florence-2 writes gendered nouns and pronouns; the polisher rewrites them to one consistent set so the trigger carries the identity instead of a description.

Read the printed samples. If the polish left artifacts, edit the regex rules and re-derive captions from the saved raw text instead of paying for the API again:

.venv/bin/python "$SKILL_DIR/scripts/recaption_dataset.py" \
  --images "./prepared_dataset/images" \
  --zip "./prepared_dataset/dataset.zip" \
  --trigger "ohwx man" \
  --pattern "subject_*.jpg" \
  --gender male \
  --reuse-report

--reuse-report re-polishes every entry in captions.json, rewrites the .txt files, and rebuilds the ZIP. It drops entries whose image no longer exists, so it stays correct after deletions.

Caption rules: references/caption-rules.md.

6. Deliver

Report:

• Kept / rejected counts
• Trigger word
• ZIP path
• 3–5 sample captions

Ready for fal-ai/flux-lora-portrait-trainer as images_data_url with trigger_phrase set to the chosen trigger.

Scripts

Script	Role
scripts/setup_env.sh	venv, pip deps, download YuNet ONNX
scripts/portrait_common.py	shared face detection, QC policy, hashing, sheets, ZIP
scripts/prepare_all_portraits.py	crop all images + split contact sheets
scripts/manual_recrop.py	apply visual crop + update manifest/sheets
scripts/qc_flags.py	framing + near-duplicate flags → qc_flags.json
scripts/recaption_dataset.py	Florence-2 captions + polish + re-zip

Degrees of freedom

• Low: 1024 crops, visual review of every contact sheet, flat ZIP pairs, [trigger] first in captions
• Medium: crop framing and automated flag thresholds
• High: manual crop corrections and whether a visibly defective source should be excluded