Kinseong/LITORA-bases

LITORA: A Large-Scale Open Portrait Relighting Dataset with AnyLight HDRI-Grounded Synthesis and Multimodal Annotations

LITORA is a 140K-pair portrait relighting dataset constructed through fully automated synthesis, combining 70K FFHQ portraits with 2,698 real-world HDRI environment maps — providing two orders of magnitude more subject diversity than any existing open relighting dataset.

Paper: LITORA: A Large-Scale Open Portrait Relighting Dataset with AnyLight HDRI-Grounded Synthesis and Multimodal Annotations (ACM MM 2026 Dataset Track)

Key Features

Property	Value
Total relighting pairs	140,036
Unique subjects (FFHQ)	70,000
HDRI environments	2,698 (AmbientCG + Poly Haven)
Resolution	1024 × 1024
VLM text annotations	100%
Per-sample quality scores	100% (6 dimensions)

Dataset Structure

Data is distributed as tar shards for efficient downloading and HuggingFace compatibility (54 shards instead of 281K individual files):

LITORA/
├── shards/
│   ├── images-00000.tar ... images-00017.tar       # 18 shards (source portraits)
│   ├── masks-00000.tar                             # 1 shard  (foreground masks)
│   ├── relit-00000.tar  ... relit-00033.tar        # 34 shards (relit targets)
│   └── backgrounds-00000.tar                       # 1 shard  (HDRI viewports)
├── metadata.jsonl
└── README.md

Shard Inventory

Shard prefix	Count	Per shard	Total	Contents
images-*	18	~5 GB	~90 GB	70K source portraits (FFHQ 1024×1024)
masks-*	1	~2.5 GB	~2.5 GB	70K foreground alpha masks
relit-*	34	~5 GB	~167 GB	140K HDRI-conditioned relit targets
backgrounds-*	1	~1.2 GB	~1.2 GB	1,349 HDRI-derived background viewports
Total	54		~260 GB

Shard ↔ Metadata Path Mapping

metadata.jsonl references original directory paths. Map them to shard prefixes as follows:

Metadata field	Path prefix	Shard prefix
image_path	images/	images-*.tar
mask_path	masks/	masks-*.tar
adv_image_path	image_adv_fbc/	relit-*.tar
bg_path	bg_env/	backgrounds-*.tar

Files are sorted by name and split into shards sequentially. Each tar stores files with their base filename only (no directory prefix).

Download Size

The full LITORA release referenced in the paper (~600 GB) comprises three components:

Component	Size	Included here	Notes
Generated data (relit targets, masks, backgrounds, metadata)	~171 GB	Yes	Novel data produced by the AnyLight pipeline
Source portraits (FFHQ 1024×1024)	~90 GB	Yes	Bundled for convenience; also available via FFHQ on HF
HDRI source maps (AmbientCG + Poly Haven)	~224 GB	No	CC0-licensed; download from source to avoid duplication

This repository: ~260 GB (generated data + source portraits — everything needed for training).

To reproduce the full pipeline or access the raw HDRI environment maps (~224 GB additional), download them from AmbientCG and Poly Haven directly, or use the download script in our pipeline repository.

Metadata Schema

Each line in metadata.jsonl is a JSON object with the following fields:

{
  "image_path": "images/00000.png",
  "mask_path": "masks/00000.png",
  "adv_image_path": "image_adv_fbc/00000__wooden_studio_02_8k__r0.png",
  "bg_path": "bg_env/wooden_studio_02_8k__r0.png",
  "hdri_id": "wooden_studio_02_8k",
  "hdri_description": "wooden studio 02 8k",
  "env_rotation": [0.157, -0.135],
  "width": 1024,
  "height": 1024,
  "text_foreground": "A baby with short dark hair wearing a ...",
  "text_background": "A dark wooden studio room with large windows ...",
  "transformation_instruction": "Apply a warm key light from the right side ...",
  "quality_metrics": {
    "dino_cosine": 0.933,
    "lpips": 0.134,
    "clip_quality": 0.498,
    "clip_score": null,
    "face_detected_src": true,
    "face_detected_relit": true,
    "landmark_shift": 0.004,
    "dino_face_cosine": 0.766,
    "boundary_gradient_ratio": 2.416,
    "boundary_score": 0.646,
    "highlight_color_sim": null,
    "composite_score": 0.731,
    "retries": 0
  },
  "gen_params": {
    "preset": "balanced",
    "steps": 25,
    "cfg": 2.0,
    "highres_scale": 1.5,
    "highres_denoise": 0.45,
    "highres_steps": 18
  },
  "seed": 42
}

Training Paradigms

LITORA supports two complementary training paradigms:

• Image-conditioned: (source portrait, mask, background viewport) → relit target
• Text-conditioned: (source portrait, transformation instruction) → relit target

Quality Metrics

All samples include per-sample quality scores across six dimensions:

1. Identity Preservation: DINOv3 cosine similarity, LPIPS perceptual distance, masked SSIM
2. Perceptual Quality: CLIP zero-shot quality scoring
3. Text–Image Alignment: CLIPScore between instruction and relit image
4. Face Integrity: MediaPipe landmark shift and face-crop DINOv3 similarity
5. Boundary Artifacts: Sobel gradient analysis along mask contour
6. Lighting Consistency: Chrominance comparison with HDRI dominant light

Composite score weights: DINO (0.20), CLIP quality (0.20), CLIP score (0.15), boundary (0.15), face (0.15), LPIPS⁻¹ (0.10), SSIM (0.05). Default gate threshold: τ = 0.55.

Quick Start

Download & Extract

# Install the HF CLI
pip install -U huggingface_hub

# Download the full dataset (~260 GB)
huggingface-cli download Kinseong/LITORA-bases --repo-type=dataset --local-dir ./LITORA

# Extract all shards into working directories
cd LITORA
mkdir -p images masks image_adv_fbc bg_env
for f in shards/images-*.tar; do tar xf "$f" -C images/; done
for f in shards/masks-*.tar; do tar xf "$f" -C masks/; done
for f in shards/relit-*.tar; do tar xf "$f" -C image_adv_fbc/; done
tar xf shards/backgrounds-00000.tar -C bg_env/

Load Metadata

import json
from pathlib import Path

with open("metadata.jsonl") as f:
    records = [json.loads(line) for line in f]

print(f"Total pairs: {len(records):,}")

# Filter by quality threshold
high_quality = [r for r in records if r["quality_metrics"]["composite_score"] > 0.7]
print(f"High-quality pairs (composite > 0.7): {len(high_quality):,}")

Selective Download

# Download only metadata (lightweight, ~200 MB)
huggingface-cli download Kinseong/LITORA-bases metadata.jsonl --repo-type=dataset

# Download only relit targets
huggingface-cli download Kinseong/LITORA-bases --repo-type=dataset --include "shards/relit-*"

# Download only source portraits
huggingface-cli download Kinseong/LITORA-bases --repo-type=dataset --include "shards/images-*"

Citation

@inproceedings{litora2026,
  title     = {LITORA: A Large-Scale Open Portrait Relighting Dataset
               with AnyLight HDRI-Grounded Synthesis and Multimodal Annotations},
  author    = {Un, Kinseong and Li, Yanfeng and Gao, Qinquan and Deng, Wei
               and Tang, Su-Kit and Chan, Ka-Hou and Tan, Tao},
  booktitle = {Proceedings of the ACM International Conference on Multimedia (MM)},
  year      = {2026}
}

License

• LITORA dataset: CC-BY-NC-SA 4.0 (compatible with FFHQ license)
• Source HDRI maps: CC0 (AmbientCG, Poly Haven)
• Pipeline code: See repository license

Ethical Considerations

This dataset contains human face images derived from FFHQ. Users should:

• Use the dataset for research purposes in accordance with the license
• Be aware of potential biases inherited from the FFHQ source distribution
• Follow responsible AI practices when training models on face data