thookham/DeOldify

By thookham · ↓ 0 · ♥ 0 · Updated 2025-12-08

What's in this repo

This page lists the files in thookham/DeOldify, including deoldify-artistic.onnx, deoldify-stable.onnx, deoldify-video.onnx.

3 ONNX models (.onnx files)
193 files, including 118 .py, 17 .md, 13 .ipynb, 7 .txt, 4 .yml, 4 .json
Last updated 2025-12-08

ONNX files

File	Size	Download
deoldify-artistic.onnx	—	Download
deoldify-stable.onnx	—	Download
deoldify-video.onnx	—	Download

How to use

These ONNX files can be loaded with ONNX Runtime. To inspect a model graph, upload the file to the Netron viewer. Download links point to the original files on Hugging Face.

README

license: mit tags:

image-colorization
gan
computer-vision
pytorch
onnx library_name: pytorch

DeOldify Model Weights

This repository contains pretrained weights for DeOldify, a deep learning model for colorizing and restoring old black and white images and videos.

Original Repository: thookham/DeOldify
Original Author: Jason Antic (jantic/DeOldify)

Model Overview

DeOldify uses a Self-Attention Generative Adversarial Network (SAGAN) with a novel NoGAN training approach to achieve stable, high-quality colorization without the typical GAN artifacts.

Three Specialized Models

Artistic - Highest quality with vibrant colors and interesting details
- Best for: General images, historical photos
- Backbone: ResNet34 U-Net
- Training: 5 NoGAN cycles, 32% ImageNet
Stable - Best for portraits and landscapes, reduced artifacts
- Best for: Faces, nature scenes
- Backbone: ResNet101 U-Net
- Training: 3 NoGAN cycles, 7% ImageNet
Video - Optimized for smooth, flicker-free video
- Best for: Video colorization, consistency
- Backbone: ResNet101 U-Net
- Training: Initial cycle only, 2.2% ImageNet

Available Files

ONNX Models (Browser/Inference)

File	Size	Description
`deoldify-art.onnx`	243 MB	Artistic model in ONNX format for browser use
`deoldify-quant.onnx`	61 MB	Quantized artistic model (75% smaller, slightly lower quality)

PyTorch Weights (Training & Inference)

Generator Weights (Main):

ColorizeArtistic_gen.pth (243 MB)
ColorizeStable_gen.pth (834 MB)
ColorizeVideo_gen.pth (834 MB)

Critic Weights (Main):

ColorizeArtistic_crit.pth (361 MB)
ColorizeStable_crit.pth (361 MB)
ColorizeVideo_crit.pth (361 MB)

PretrainOnly Weights (For continued training):

ColorizeArtistic_PretrainOnly_gen.pth (729 MB)
ColorizeArtistic_PretrainOnly_crit.pth (1.05 GB)
ColorizeStable_PretrainOnly_crit.pth (1.05 GB)
ColorizeVideo_PretrainOnly_crit.pth (1.05 GB)

Note: Stable and Video PretrainOnly generators are split files hosted on GitHub Releases.

Usage

Browser (ONNX)

<!DOCTYPE html>
<html>
<head>
  <script src="https://cdn.jsdelivr.net/npm/onnxruntime-web/dist/ort.min.js"></script>
</head>
<body>
  <script>
    async function colorize() {
      // Load model from Hugging Face
      const session = await ort.InferenceSession.create(
        "https://huggingface.co/thookham/DeOldify/resolve/main/deoldify-art.onnx"
      );
      
      // Run inference (see full example in GitHub repo)
      // ...
    }
  </script>
</body>
</html>

PyTorch (Python)

from huggingface_hub import hf_hub_download
import torch

# Download model weights
model_path = hf_hub_download(
    repo_id="thookham/DeOldify",
    filename="ColorizeArtistic_gen.pth"
)

# Load weights (requires deoldify package installed)
# See GitHub repository for full usage examples

Installation

# Clone the main repository
git clone https://github.com/thookham/DeOldify
cd DeOldify

# Install dependencies
pip install -r requirements.txt

# Download a model
from huggingface_hub import hf_hub_download
model = hf_hub_download(repo_id="thookham/DeOldify", filename="ColorizeStable_gen.pth")

Technical Details

Architecture

Generator: U-Net with ResNet34/101 backbone, spectral normalization, self-attention layers
Critic: PatchGAN discriminator
Loss: Perceptual loss (VGG16) + GAN loss

NoGAN Training

A novel training approach that combines:

Generator pretraining with feature loss
Critic pretraining on generated images
Short GAN training (30-60 minutes) at inflection point
Optional cycle repeats for more colorful results

This eliminates typical GAN artifacts while maintaining realistic colorization.

Training Data

Dataset: ImageNet subsets (1-32% depending on model)
Resolution: 192px during training
Augmentation: Gaussian noise for video stability

Model Card

Model Details

Developed by: Jason Antic (original), Travis Hookham (modernization)
Model type: Conditional GAN for image-to-image translation
Language(s): N/A (computer vision)
License: MIT
Parent Model: Based on FastAI U-Net and Self-Attention GAN papers

Intended Use

Primary Use: Colorizing black and white photographs and videos
Out-of-Scope: Real-time processing, guaranteed historical accuracy

Limitations

Colors may not be historically accurate
Performance degrades on very low quality/damaged images
Artistic model may require render_factor tuning
Video model trades some color vibrancy for consistency

Related Models & Resources

Similar Colorization Models on Hugging Face

GAN-based Colorization:

Hammad712/GAN-Colorization-Model - GAN model for grayscale to color transformation
jessicanono/filparty_colorization - ResNet-based model for historical photos

Stable Diffusion-based:

rsortino/ColorizeNet - ControlNet adaptation of SD 2.1 for colorization
AlekseyCalvin/ColorizeTruer_KontextFluxVar6_BySAP - Advanced Flux-based colorization

Interactive Demos (Spaces):

Why Choose DeOldify?

DeOldify stands out for:

NoGAN Training: Unique approach eliminating typical GAN artifacts
Specialized Models: Three purpose-built models (Artistic, Stable, Video)
Video Support: Flicker-free temporal consistency
Proven Track Record: Powers MyHeritage InColor and widely adopted
ONNX Support: Browser-ready models for offline use

Citation

If you use these models, please cite:

@misc{deoldify,
  author = {Antic, Jason},
  title = {DeOldify},
  year = {2019},
  publisher = {GitHub},
  url = {https://github.com/jantic/DeOldify}
}

License

MIT License. See LICENSE file.

Model info

pytorch

pytorch onnx image-colorization gan computer-vision license:mit region:us

↓ 0 ♥ 0

Created: 2025-12-01

Updated: 2025-12-08

View on Hugging Face

Files (193)

.gitattributes

.github/CODEOWNERS

.github/ISSUE_TEMPLATE/bug_report.md

.github/ISSUE_TEMPLATE/feature_request.md

.github/PULL_REQUEST_TEMPLATE.md

.gitignore

.pre-commit-config.yaml

.pylintrc

.travis.yml

CHANGELOG.md

CODE_OF_CONDUCT.md

CONTRIBUTING.md

ColorFIDBenchmarkArtistic.ipynb

ColorizeTrainingArtistic.ipynb

ColorizeTrainingStable.ipynb

ColorizeTrainingStableLargeBatch.ipynb

ColorizeTrainingVideo.ipynb

ColorizeTrainingWandb.ipynb

HF_README.md

ImageColorizer.ipynb

ImageColorizerArtisticTests.ipynb

ImageColorizerColab.ipynb

ImageColorizerColabStable.ipynb

ImageColorizerStableTests.ipynb

LICENSE

MANIFEST.in

README.md

ROADMAP.md

SECURITY.md

TODO.md

USAGE_EXAMPLES.md

VideoColorizer.ipynb

VideoColorizerColab.ipynb

akbartus_tree.json

assets.txt

browser/README.md

browser/artistic.html

browser/index.html

browser/quantized.html

browser/serve.ps1

check_remote_models.py

check_remote_models_v2.py

compare_models.py

deoldify-artistic.onnx ONNX

deoldify-artistic.onnx.data

deoldify-stable.onnx ONNX

deoldify-stable.onnx.data

deoldify-video.onnx ONNX

deoldify-video.onnx.data

deoldify/__init__.py

deoldify/_device.py

deoldify/augs.py

deoldify/critics.py

deoldify/dataset.py

deoldify/device_id.py

deoldify/fastai_compat.py

deoldify/filters.py

deoldify/generators.py

deoldify/layers.py

deoldify/loss.py

deoldify/save.py

deoldify/unet.py

deoldify/visualize.py

docs/DEPLOYMENT_GUIDE.md

docs/HARDWARE_GUIDE.md

docs/intel_gpu_setup.md

docs/nvidia_setup.md

environment-legacy.yml

environment.yml

environment_intel.yml

fastai/LICENSE

fastai/__init__.py

fastai/basic_data.py

fastai/basic_train.py

fastai/basics.py

fastai/callback.py

fastai/callbacks/__init__.py

fastai/callbacks/csv_logger.py

fastai/callbacks/fp16.py

fastai/callbacks/general_sched.py

fastai/callbacks/hooks.py

fastai/callbacks/loss_metrics.py

fastai/callbacks/lr_finder.py

fastai/callbacks/mem.py

fastai/callbacks/misc.py

fastai/callbacks/mixup.py

fastai/callbacks/mlflow.py

fastai/callbacks/one_cycle.py

fastai/callbacks/oversampling.py

fastai/callbacks/rnn.py

fastai/callbacks/tensorboard.py

fastai/callbacks/tracker.py

fastai/collab.py

fastai/core.py

fastai/data_block.py

fastai/datasets.py

fastai/distributed.py

fastai/gen_doc/__init__.py

fastai/gen_doc/autogen.tpl

fastai/gen_doc/convert2html.py

fastai/gen_doc/core.py

fastai/gen_doc/docstrings.py

fastai/gen_doc/doctest.py

fastai/gen_doc/gen_notebooks.py

fastai/gen_doc/hide.tpl

fastai/gen_doc/jekyll.tpl

fastai/gen_doc/nbdoc.py

fastai/gen_doc/nbtest.py

fastai/general_optimizer.py

fastai/imports/__init__.py

fastai/imports/core.py

fastai/imports/torch.py

fastai/launch.py

fastai/layers.py

fastai/metrics.py

fastai/script.py

fastai/sixel.py

fastai/tabular/__init__.py

fastai/tabular/data.py

fastai/tabular/models.py

fastai/tabular/transform.py

fastai/test_registry.json

fastai/text/__init__.py

fastai/text/data.py

fastai/text/interpret.py

fastai/text/learner.py

fastai/text/models/__init__.py

fastai/text/models/awd_lstm.py

fastai/text/models/bwd_forget_mult_cuda.cpp

fastai/text/models/bwd_forget_mult_cuda_kernel.cu

fastai/text/models/forget_mult_cuda.cpp

fastai/text/models/forget_mult_cuda_kernel.cu

fastai/text/models/qrnn.py

fastai/text/models/transformer.py

fastai/text/transform.py

fastai/torch_core.py

fastai/train.py

fastai/utils/__init__.py

fastai/utils/check_perf.py

fastai/utils/collect_env.py

fastai/utils/ipython.py

fastai/utils/mem.py

fastai/utils/mod_display.py

fastai/utils/pynvml_gate.py

fastai/utils/show_install.py

fastai/version.py

fastai/vision/__init__.py

fastai/vision/cyclegan.py

fastai/vision/data.py

fastai/vision/gan.py

fastai/vision/image.py

fastai/vision/interpret.py

fastai/vision/learner.py

fastai/vision/models/__init__.py

fastai/vision/models/cadene_models.py

fastai/vision/models/darknet.py

fastai/vision/models/presnet.py

fastai/vision/models/unet.py

fastai/vision/models/wrn.py

fastai/vision/models/xception.py

fastai/vision/models/xresnet.py

fastai/vision/models/xresnet2.py

fastai/vision/transform.py

fastai/vision/tta.py

fastai/widgets/__init__.py

fastai/widgets/class_confusion.py

fastai/widgets/image_cleaner.py

fastai/widgets/image_downloader.py

fid/LICENSE

fid/fid_score.py

fid/inception.py

found_urls.txt

hashes.txt

models.json

models/.gitkeep

reassemble_models.py

requirements-colab.txt

requirements-dev.txt

requirements.txt

requirements_intel.txt

resource_images/twitter.svg

resource_images/watermark.png

scripts/download_models.py

setup.py

tests/test_device.py

tests/test_fastai_compat.py

thookham_tree.json

tox.ini

upload_to_hf.py

verify_deflicker.py

verify_deflicker_static.py

verify_models.py

verify_refactor.py

Related models

More models

No task type declared.