bridenmj/wound-stylegan

Wound StyleGAN Generator

This repository contains the trained StyleGAN generator used in the Wound Forecasting project.

The generator was trained on longitudinal porcine wound images and was used as the image-representation bottleneck for e4e inversion and DyneODE latent-space forecasting.

This repository is currently private while public-release authority and licensing are being confirmed.

Released artifact

Field	Value
File	network-snapshot-005000_jmir_50000_redo_.pt
Checkpoint generator field	g_ema
Additional field	latent_avg
Output resolution	256 x 256
Style dimension	512
Mapping depth	8
Channel multiplier	2

The checkpoint is a converted PyTorch state-dictionary package rather than an original StyleGAN pickle.

Loading architecture

The final DyneODE implementation constructs the generator with:

from stylegan2.model import Generator
import torch

generator = Generator(
    256,
    512,
    8,
    channel_multiplier=2,
)

checkpoint = torch.load(
    checkpoint_path,
    map_location="cpu",
    weights_only=False,
)

generator.load_state_dict(
    checkpoint["g_ema"],
    strict=True,
)

generator.eval()

The final paper notebook verified that the checkpoint loaded with zero missing and zero unexpected generator parameters.

Provenance

The generator was trained using the Collapse by Conditioning / transitional-cGAN implementation:

• https://github.com/mshahbazi72/transitional-cGAN
• https://arxiv.org/abs/2201.06578

That implementation is based on NVIDIA StyleGAN2-ADA:

• https://github.com/NVlabs/stylegan2-ada-pytorch

LICENSE_ADA.txt is included as the upstream notice. Its inclusion documents upstream provenance and does not assign a new project license to this repository.

Project resources

• Project code: https://github.com/bridenmj/wound-forecasting
• Processed dataset: https://huggingface.co/datasets/bridenmj/porcine-wound-forecasting-processed

Reproducibility note

The recovered historical Kubernetes launcher used four GPUs and invoked train_4gpu.sh. The complete historical train.py argument list was not embedded in the final DyneODE notebook. The exact released generator artifact and its inference architecture are preserved here.

Intended use

This generator is provided for research reproduction and investigation of longitudinal wound-image representations.

Project license

The original project-trained weights released in this repository are licensed under the Creative Commons Attribution–NonCommercial 4.0 International license (CC BY-NC 4.0).

Third-party implementations, base models, tokenizers, and other external dependencies are not relicensed by this repository and remain subject to their respective licenses and terms. Any upstream license or notice files included in this repository continue to apply to the corresponding upstream materials.