Skip to content
Get packs

Bringing Old Photos Back to Life

Research code that restores old, damaged, or scratched photos using a deep learning model.

GitHub Download v1.0
Stars
15.7k
Forks
2.1k
Licence
MIT
Language
Python
Last update
3.0 years ago
Type
AI model / agent

This repository is archived and no longer maintained.

What it is

This is the official PyTorch implementation of a research project from CVPR 2020 that repairs old photos, removing scratches, fixing faded or degraded quality, and enhancing faces. It runs a pipeline of pretrained neural networks over a folder of images and saves the restored results. The repository is archived, so it is no longer actively maintained.

  • Restore a folder of scanned or faded old photos with one run.py command
  • Restore photos that have scratches by using the --with_scratch option
  • Run the high-resolution option (--HR) on scratched images
  • Run scratch detection to generate masks for your own damaged photos
  • Experiment with the GUI to remove scratches from a single image

Great for

  • Researchers studying image restoration with GANs
  • Developers who want to run a pretrained photo restoration pipeline on an NVIDIA GPU
  • Students learning how a multi-stage restoration pipeline is built and trained
  • People comfortable with Python setup and command-line tools

Not ideal for

  • Non-technical users wanting a polished app, since the README describes the code as research-oriented and not optimized for product use
  • Users without an NVIDIA GPU, since the README says the code is tested on Ubuntu with Nvidia GPUs and CUDA (CPU mode is mentioned only as a flag)
  • Anyone needing high-resolution or arbitrary-size restoration, since the README says the model was pretrained on 256x256 images and may not work well at other resolutions
  • Teams needing ongoing support or updates, since the repository is archived

How to get it

  1. 1

    Clone and build from source

    You'll need: Python 3.6 or newer · Nvidia GPU with CUDA installed (code tested on Ubuntu) · Git · wget and unzip for downloading pretrained checkpoints

    Download the pretrained face checkpoints

    wget https://github.com/microsoft/Bringing-Old-Photos-Back-to-Life/releases/download/v1.0/face_checkpoints.zip

    Download the pretrained global checkpoints

    wget https://github.com/microsoft/Bringing-Old-Photos-Back-to-Life/releases/download/v1.0/global_checkpoints.zip

    Install Python dependencies

    pip install -r requirements.txt
  2. 2

    Run it

    • Clone the repository and complete the Synchronized-BatchNorm and dlib landmark model setup described in the README
    • Download and unzip the face and global checkpoints into Face_Enhancement and Global
    • Install dependencies with pip install -r requirements.txt
    • Run python run.py --input_folder [test_image_folder_path] --output_folder [output_path] --GPU 0 using absolute paths
    • Find results in ./output_path/final_output/
  3. 3

    Get the most out of it

    • Add --with_scratch for photos that have visible scratches
    • Add --HR together with --with_scratch for high-resolution scratched images
    • Check intermediate outputs in the output folder to see results from each step
    • Use the GUI.py file for single-image scratch removal
    • Use the Colab demo linked in the README if you lack a local GPU

Downloads

Source code (.zip)master branch · straight from GitHubRelease v1.02 files incl. face_checkpoints.zipOfficial website & docsarxiv.org/abs/2004.09484

Strengths

  • Official implementation of a peer-reviewed CVPR 2020 oral paper with a published pretrained model
  • Single command runs the full pipeline for images with or without scratches
  • MIT license permits reuse and modification

Trade-offs

  • Archived repository with last push in October 2023, so no future fixes are expected
  • Research-grade code: the README says it has not been optimized for running performance or product use, and 256x256 pretrained resolution limits results on other sizes

How it compares

ProjectStarsTypeLevelUpdatedBest for
Bringing-Old-Photos-Back-to-Life (this)15.7kAI model / agentadvanced3.0 years agoResearchers studying image restoration with GANs
DeOldify18.5kAI model / agentintermediate2.0 years agoPeople who want to colorize old family or archival photos without paying for a service
JARVIS25.5kAI model / agentadvanced1.3 years agoResearchers studying LLM-driven task planning and tool use
FinGPT21.5kAI model / agentadvanced12 days agoDevelopers who want to experiment with finance-tuned LLMs for sentiment analysis
LaTeX-OCR16.6kAI model / agentintermediate1.8 years agoStudents who want to turn a handwritten or printed formula photo into editable LaTeX
annotated_deep_learning_paper_implementations67.6kLearning resourceintermediate4 days agoStudents learning how transformers, GANs, or optimizers work internally

Stats synced from GitHub. The assessment above was drafted with AI from the project's README on 2026-10-29 and may miss recent changes — check the README before relying on a command. MIT licence: free to use, modify and distribute, including commercially, as long as you keep the copyright notice.