DeOldify adds inferred color to monochrome images. The useful result is a careful interpretation of an existing photograph, with its people and details intact. That goal differs from generating a new scene, removing scratches or inventing a sharper face.
Start with the original archived project
Jason Antic's repository was archived on October 19, 2024 and is read-only. Its README links to notebooks and alternative interfaces, but their current runtime availability needs a separate check. An old tutorial opening successfully does not mean its dependencies or model download still work today.
The repository states that its code and listed pretrained weights use the MIT license. That does not grant rights to a photograph you downloaded elsewhere. Record the image's source and permission separately from the software license.
Choose Artistic, Stable or Video for the task
The README describes three models with different emphases. Use that distinction to choose a starting point; it is the author's description rather than our benchmark. Do not assume a third-party service named DeOldify exposes all three or the same settings.
| Model | Documented emphasis | First review to make |
|---|---|---|
| Artistic | More vibrant image colorization with tuning | Color spill and overly strong skin or fabric color |
| Stable | Portraits and landscapes with restrained coloration | Uneven faces, gray patches and muted detail |
| Video | More consistent color across moving footage | Changes in color between adjacent frames |
Prepare a photograph without overwriting the source
Keep the original scan untouched and make a working copy. If you are scanning a print, capture the full photograph with even lighting and enough detail for the intended display size. Crop only the working copy. Keep handwritten notes or an envelope with the source record rather than baking them into the colorization input.
Choose an image with several inspectable regions: a face or clearly visible object, neutral fabric, foliage or architecture and a bright area. A train-station photograph is a useful exercise because rails, luggage and station walls expose different kinds of color spill. Use an image you own or one with documented permission.
Our cover depicts an invented station photograph in monochrome and an imagined color treatment. It is original editorial artwork, not a DeOldify before-and-after test or evidence of the station's real colors.
Make a controlled colorization comparison
- Open the chosen notebook or interface and confirm its model download and runtime before uploading the full archive.
- Load one working image and record the model, settings and output dimensions.
- Generate a baseline with the interface's recommended settings.
- Change one exposed setting, such as render factor, and save the second result separately.
- Compare the same regions at the same display size before choosing a treatment.
The original README discusses render factor as a rendering-resolution control, including stability tradeoffs. An exposed control's useful range depends on your implementation and hardware. Do not blindly copy a value from a different notebook or confuse it with the final export's pixel dimensions.
Download the original archive colorization review sheet
Inspect inferred color instead of calling it recovered truth
Look for color crossing an object's boundary, inconsistent treatment of the same material and remaining gray patches on a subject. Compare skin, signs, uniforms and vehicle paint against independent evidence if the project requires historical accuracy. A convincing green train can still be the wrong green train.
Keep a short decision note: which regions are plausible, which are uncertain and which were corrected manually. If you publish the interpretation, describe it as colorized and preserve access to the monochrome version. This is especially useful in a narrated history video where the audience might otherwise treat an invented color as documentary evidence.
For video, inspect temporal consistency
Test a short section before processing a long film. Pick a section with movement and a lighting change, not only a static frame. Watch for clothing that changes color when a person turns and patches that flicker between nearby frames. Inspect the ordinary-speed playback as well as individual frames.
Keep audio and editing decisions separate from the colorization step. Check that your finished export preserves the intended duration and synchronization. A folder of correctly colored frames is not yet a reviewed video deliverable.
Choose an alternative when the task changes
Use the image upscaling guide when the job is resolution and detail, or the photo deblur guide when blur is the main problem. For newly invented creator imagery, use an image-generation workflow rather than describing it as archive restoration. Do not use a generative replacement face as proof of what the original person looked like.
