deepmancer/bald_konverter

Bald Converter LoRA Weights

Bald Converter is the LoRA-based in-context bald-generation module used by HairPort: In-context 3D-Aware Hair Import and Transfer for Images, accepted to ACM SIGGRAPH 2026.

This repository contains two FLUX.1 Kontext LoRA adapters for generating realistic bald versions of portrait images while preserving identity, facial structure, clothing, background, lighting, and camera style.

What This Repository Contains

File	Size	Variant	Purpose
bald_konvertor_wo_seg_000003400.safetensors	171,969,368 bytes	wo_seg	Fast two-panel bald conversion without segmentation preprocessing
bald_konvertor_w_seg_000004900.safetensors	171,969,368 bytes	w_seg	Segmentation-guided four-panel refinement for higher-quality conversion

The filename uses konvertor for historical compatibility with the released checkpoint names. The model/component name in the paper and codebase is Bald Converter or BaldKonverter.

Model Details

Item	Value
Base model	black-forest-labs/FLUX.1-Kontext-dev
Adapter type	LoRA
Framework	Diffusers
Weight format	Safetensors
Training data	deepmancer/baldy
Primary task	Portrait bald conversion / hair removal
Released with	HairPort, ACM SIGGRAPH 2026

These files are adapter weights only. They do not include the base FLUX model, preprocessing models, or the HairPort inference code.

Modes

Mode	Checkpoint	Speed	Quality	Preprocessing
wo_seg	bald_konvertor_wo_seg_000003400.safetensors	Fast	Good	None
w_seg	bald_konvertor_w_seg_000004900.safetensors	Slower	Best	SAM3, BEN2, and face parsing
auto	Uses both checkpoints	Slower	Best	Runs wo_seg, then refines with w_seg

wo_seg assembles a two-panel image and asks FLUX to generate the bald version in the right panel. w_seg uses a four-panel layout with segmentation context to improve structure preservation and remove residual hair more reliably.

Installation

The recommended way to use these weights is through the HairPort source checkout:

git clone https://github.com/deepmancer/HairPort.git
cd HairPort
bash scripts/install.sh

If the base FLUX model or related gated assets require authentication, log in to Hugging Face before running inference:

huggingface-cli login

The LoRA checkpoints are downloaded automatically from this repository on first use.

Quick Start

Run Bald Converter from the HairPort repository root:

python -m hairport.bald_konverter.cli \
  --input photo.jpg \
  --output bald.png \
  --mode auto

Fast mode without segmentation preprocessing:

python -m hairport.bald_konverter.cli \
  --input photo.jpg \
  --output bald_fast.png \
  --mode wo_seg

Batch processing:

python -m hairport.bald_konverter.cli \
  --input-dir ./faces \
  --output-dir ./bald \
  --mode auto

Save intermediate masks and FLUX input grids:

python -m hairport.bald_konverter.cli \
  --input photo.jpg \
  --output bald.png \
  --mode auto \
  --save-intermediates

Python API

from hairport.bald_konverter import BaldKonverterPipeline

pipeline = BaldKonverterPipeline(mode="auto", device="cuda")
result = pipeline("portrait.jpg", return_intermediates=True)

result.bald_image.save("bald.png")

# Optional debugging / visualization outputs
if result.flux_input_wo_seg is not None:
    result.flux_input_wo_seg.save("flux_input_wo_seg.png")
if result.flux_input_w_seg is not None:
    result.flux_input_w_seg.save("flux_input_w_seg.png")

pipeline.teardown()

Use a custom local checkpoint path if needed:

from hairport.bald_konverter import BaldKonverterPipeline

pipeline = BaldKonverterPipeline(
    mode="wo_seg",
    lora_path_wo_seg="/path/to/bald_konvertor_wo_seg_000003400.safetensors",
)
result = pipeline("portrait.jpg")
result.bald_image.save("bald.png")

Download Checkpoints Manually

from huggingface_hub import hf_hub_download

repo_id = "deepmancer/bald_konverter"

wo_seg_path = hf_hub_download(
    repo_id=repo_id,
    filename="bald_konvertor_wo_seg_000003400.safetensors",
)

w_seg_path = hf_hub_download(
    repo_id=repo_id,
    filename="bald_konvertor_w_seg_000004900.safetensors",
)

print(wo_seg_path)
print(w_seg_path)

Requirements

For inference through HairPort:

• Python >= 3.10
• CUDA GPU recommended
• Approximately 24 GB VRAM recommended for FLUX inference
• torch, diffusers, transformers, safetensors, and huggingface-hub
• For w_seg / auto: SAM3, BEN2, and face-parsing dependencies
• Optional FLAME/SHeaP assets for FLAME-guided segmentation

Refer to the HairPort README for the current full dependency recipe.

Training Data

The Bald Converter LoRAs were trained with Baldy, a 6,400-sample synthetic paired dataset of hair/bald portraits generated for HairPort.

• Dataset: deepmancer/baldy
• Each sample includes paired hair_image and bald_image columns, plus hair renders, background images, camera metadata, and rendering parameters.

Intended Use

These weights are intended for research and development in:

• bald conversion and hair removal
• hairstyle-transfer preprocessing
• 3D-aware hair import and transfer systems
• paired image-to-image generation experiments
• HairPort-style source-face preparation

Limitations And Responsible Use

Bald Converter is designed for portrait preprocessing and research use. Outputs may contain artifacts, incomplete hair removal, identity drift, or failures under unusual poses, occlusions, lighting, accessories, or image crops.

The model should not be used to misrepresent people, produce deceptive identity edits, or infer sensitive attributes. Users should review generated outputs before downstream use.

Related Resources

• Project page: deepmancer.github.io/HairPort
• Code: github.com/deepmancer/HairPort
• Baldy dataset: deepmancer/baldy
• Weights: deepmancer/bald_konverter

Citation

If you use Bald Converter, 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 = {ACM SIGGRAPH 2026},
  year      = {2026}
}

License

The LoRA weights in this repository are released under the Apache License 2.0. Use of the base model and any external preprocessing models is governed by their respective licenses and terms.