Baldy Dataset

Baldy is a synthetic paired image dataset for bald conversion, hairstyle-transfer preprocessing, and 3D-aware hair research. It contains 6,400 identity-consistent hair/bald image pairs with auxiliary hair renders, background images, camera parameters, and rendering metadata.

Baldy is released with HairPort: In-context 3D-aware Hair Import and Transfer for Images, an ACM SIGGRAPH 2026 paper. The paper is available on arXiv (arXiv:2606.12562), and the ACM DOI is 10.1145/3799902.3811046.

Overview

Each Baldy sample provides a photorealistic image of a person with hair and a corresponding bald version of the same subject. The paired structure is designed to support training and evaluation of bald-conversion models, including the Bald Converter used by HairPort.

The dataset also includes intermediate rendering assets and metadata, making it useful for controlled experiments that require camera/view information, hairstyle provenance, or synthetic render supervision.

What's Included

Component	Description
hair_image	Photorealistic image of the subject with hair
bald_image	Photorealistic bald version of the same subject
hair_render	Blender-rendered hairstyle on a transparent background
bald_render	Blender-rendered bald hairline layer on a transparent background
background_image	Generated or rendered scene background
Camera metadata	Focal length, location, and rotation in Blender coordinates
Appearance metadata	Hair material values, lighting preset, expression, body metadata
Source metadata	Hairstyle source dataset and source-specific hairstyle ID

Dataset Construction

Baldy was generated with a multi-stage synthetic data pipeline:

1. 3D hairstyle preparation. Hairstyles from DiffLocks, Hair20K, USC-HairSalon, and CT2Hair are aligned to SMPL-X head/body configurations with pose, expression, and garment variation.
2. Blender rendering. Hair, body, camera, lighting, and material parameters are rendered at 1024 x 1024 resolution.
3. Photorealistic paired generation. ControlNet++ and SDXL-based generation, followed by FLUX Kontext refinement, are used to produce identity-consistent hair/bald image pairs.

Dataset Statistics

Split	Samples
train	6,400

View Distribution

View	Samples
front	6,009
side	292
back	99

The current release is front-view dominant. Users training view-balanced models may want to account for this distribution during sampling or evaluation.

Hairstyle Source Distribution

Source	Samples
difflocks	3,197
hair20k	2,824
usc	370
ct2hair	9

Data Fields

Identity, View, And Source

Field	Type	Description
hairstyle_id	string	Unique zero-padded sequential ID, such as "000042"
view	string	Camera view: "front", "side", or "back"
source	string	Hairstyle source: "difflocks", "hair20k", "usc", or "ct2hair"
hairstyle_source_id	string	Source-specific hairstyle identifier
views_available	string	Pipe-separated list of views available for the hairstyle, such as `"front

Image Columns

Field	Type	Description
hair_image	Image	Photorealistic image of the subject with hair, decoded as a PIL image
bald_image	Image	Photorealistic bald image of the same subject, decoded as a PIL image
hair_render	Image	Blender-rendered hair layer, decoded as a PIL image
bald_render	Image	Blender-rendered bald hairline layer, decoded as a PIL image
background_image	Image	Background image, decoded as a PIL image

Camera And Rendering Metadata

Field	Type	Description
render_params_json	string	Full Blender render parameters as an embedded JSON string
background_prompt	string	Text prompt used to generate the background, empty when unavailable
camera_focal_length	float64	Camera focal length in millimeters
camera_location_x	float64	Camera X position in Blender world coordinates
camera_location_y	float64	Camera Y position in Blender world coordinates
camera_location_z	float64	Camera Z position in Blender world coordinates
camera_rotation_x	float64	Camera X rotation in radians
camera_rotation_y	float64	Camera Y rotation in radians
camera_rotation_z	float64	Camera Z rotation in radians
lighting_preset	string	Lighting preset name

Appearance Metadata

Field	Type	Description
body_gender	string	SMPL-X body gender configuration
face_expression	string	Facial expression label; empty for some samples
hair_melanin	float64	Hair melanin value controlling color darkness
hair_roughness	float64	Hair surface roughness value
has_garments	bool	Whether BEDLAM clothing is applied to the body

Quick Start

Install the Hugging Face datasets package if needed:

pip install datasets

Load the dataset:

from datasets import load_dataset

ds = load_dataset("deepmancer/baldy", split="train")
sample = ds[0]

hair = sample["hair_image"]  # PIL.Image: subject with hair
bald = sample["bald_image"]  # PIL.Image: same subject without hair

Save a paired sample:

sample["hair_image"].save("hair.png")
sample["bald_image"].save("bald.png")
sample["hair_render"].save("hair_render.png")
sample["bald_render"].save("bald_render.png")
sample["background_image"].save("background.png")

Common Usage Patterns

Filter by camera view:

front = ds.filter(lambda row: row["view"] == "front")
side = ds.filter(lambda row: row["view"] == "side")
back = ds.filter(lambda row: row["view"] == "back")

Filter by hairstyle source:

hair20k = ds.filter(lambda row: row["source"] == "hair20k")
difflocks = ds.filter(lambda row: row["source"] == "difflocks")

Read Blender render parameters:

import json

params = json.loads(sample["render_params_json"])
print(params.keys())

Stream the dataset without downloading all shards:

from datasets import load_dataset

stream = load_dataset("deepmancer/baldy", split="train", streaming=True)
for row in stream:
    print(row["hairstyle_id"], row["view"], row["source"])
    break

File Format

Baldy is stored as sharded Parquet files with embedded image bytes:

data/
├── train-00000-of-NNNNN.parquet
├── train-00001-of-NNNNN.parquet
├── ...
└── train-NNNNN-of-NNNNN.parquet

No external image files are required. Image columns are decoded automatically as PIL images by the datasets library.

Intended Uses

Baldy is intended for research and development in:

• bald-conversion model training
• hairstyle-transfer preprocessing
• paired image-to-image translation
• synthetic data studies for hair and head rendering
• controlled evaluation of hair removal and reconstruction pipelines
• HairPort-style 3D-aware hair import and transfer systems

Limitations And Responsible Use

Baldy is a synthetic dataset. Its distribution reflects the hairstyle sources, SMPL-X configurations, rendering settings, and generative refinement pipeline used to create it. The current release is also front-view dominant.

Generated samples may contain artifacts or biases inherited from the rendering and image-generation stages. Users should inspect samples before using the dataset in production settings.

This dataset is not designed as a demographic benchmark and should not be used for sensitive identity, demographic, or attribute-inference tasks.

Related Resources

• Paper (arXiv): arXiv:2606.12562
• Project page: deepmancer.github.io/HairPort
• Code: github.com/deepmancer/HairPort
• Bald Converter LoRA weights: deepmancer/bald_konverter
• Baldy dataset: deepmancer/baldy

Citation

If you use Baldy or HairPort in your research, please cite:

@inproceedings{heidari2026hairport,
  title     = {HairPort: In-context 3D-aware Hair Import and Transfer for Images},
  author    = {A. Heidari and A. Alimohammadi and W. Michel Pinto Lira and A. Bar-Lev and A. Mahdavi-Amiri},
  booktitle = {Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Papers (SIGGRAPH Conference Papers '26)},
  year      = {2026},
  isbn      = {979-8-4007-2554-8/2026/07},
  doi       = {10.1145/3799902.3811046},
  url       = {https://doi.org/10.1145/3799902.3811046},
  location  = {Los Angeles, CA, USA}
}

License

Baldy is released under the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License (CC BY-NC-ND 4.0).

Users are responsible for complying with the licenses and terms of any upstream tools, models, or source assets they use alongside this dataset.