You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

By requesting access to these weights you agree to use them for non-commercial research only, and to cite the AHS paper (arxiv:2604.15857) in any derivative work.

AHS: Adaptive Head Synthesis via Synthetic Data Augmentations

Official pretrained weights for AHS.

Inference code: https://github.com/KEH0T0/AHS (branch: inference)
Paper: https://arxiv.org/abs/2604.15857

Repo contents

path	size	purpose
`id_encoder_weights.pth`	1.6 GB	PhotoMaker-style ID encoder (`photomaker_encoder.IDEncoder`)
`graphonomy_weights.pth`	166 MB	Graphonomy DeepLab-Xception human parsing net (for head masks)
`checkpoint-179160/`	19 GB	SDXL fine-tune (UNet + tryon UNet-encoder + text encoders + VAE + tokenizers)

checkpoint-179160/ is a diffusers-format directory and can be loaded via from_pretrained subfolder calls (see infer_xl_base.py).

Quick download (after gate approval)

pip install huggingface_hub
huggingface-cli login   # token with 'read' scope

# All weights
huggingface-cli download Keh0t0/AHS-weights --local-dir ./AHS-weights

# Or only what you need:
huggingface-cli download Keh0t0/AHS-weights \
    id_encoder_weights.pth graphonomy_weights.pth \
    --local-dir weights/

huggingface-cli download Keh0t0/AHS-weights \
    --include "checkpoint-179160/*" --local-dir checkpoints/

Usage

These weights are consumed by the inference code at github.com/KEH0T0/AHS. The README there walks through environment setup, preprocessing (GAGAvatar_track head alignment, Graphonomy head mask, DensePose, EMOCA normal swap), and inference.

Minimal quick-run over the 3-pair sample dataset shipped with the code:

git clone https://github.com/KEH0T0/AHS.git -b inference
cd AHS
conda env create -f environment.yaml && conda activate ahs

# Link the downloaded weights
mkdir -p weights
ln -sf /absolute/path/to/AHS-weights/id_encoder_weights.pth  weights/id_encoder_weights.pth
ln -sf /absolute/path/to/AHS-weights/graphonomy_weights.pth  weights/graphonomy_weights.pth

BASE_MODEL=/absolute/path/to/AHS-weights/checkpoint-179160 \
    CUDA_VISIBLE_DEVICES=0 bash infer_sample.sh

Output lands in ./sample_result/final/<head>_<body>.png (triplet format: [head | body | swapped]).

Method summary

AHS takes a (body image, head reference) pair and runs two-pass SDXL inpainting over the head region of the body. Identity is transferred via (a) PhotoMaker-style ID encoder that reads a face crop of the head reference and injects it at the img trigger token, and (b) IP-Adapter style CLIP image conditioning from the GAGAvatar-aligned head image. Shape guidance comes from a merged map of body DensePose (below the neck) and EMOCA face-normals of the head rendered in the body's pose (above the neck).

License

Released under CC BY-NC 4.0 for research use. Do not use these weights for commercial purposes. Upstream components follow their own licenses:

SDXL base model — CreativeML Open RAIL++-M
IDM-VTON — see https://github.com/yisol/IDM-VTON
Graphonomy — see https://github.com/Gaoyiminggithub/Graphonomy

Citation

@misc{kang2026ahsadaptiveheadsynthesis,
      title={AHS: Adaptive Head Synthesis via Synthetic Data Augmentations},
      author={Taewoong Kang and Hyojin Jang and Sohyun Jeong and Seunggi Moon and Gihwi Kim and Hoon Jin Jung and Jaegul Choo},
      year={2026},
      eprint={2604.15857},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2604.15857},
}

Acknowledgements

IDM-VTON — base UNet / tryon pipeline
IP-Adapter — image prompt adapter
PhotoMaker — ID encoder design
Graphonomy — human parsing
EMOCA — 3D face normals
GAGAvatar_track — head alignment
DensePose — body pose

Downloads last month: -

Paper for Keh0t0/AHS-weights

AHS: Adaptive Head Synthesis via Synthetic Data Augmentations

Paper • 2604.15857 • Published 6 days ago