This colorization pioneer finally faded to black

What it does

DeOldify is a deep-learning project that colorizes and restores old images and silent film footage. It provides separate image models—one labeled “artistic” that favors vivid color, and a “stable” variant that suppresses glitches—along with a video pipeline that renders individual frames without any temporal modeling. The repository is archived as of October 2024; the author considers it unmaintainable and “ancient” by modern deep-learning standards.

The interesting bit

The project’s core idea is NoGAN training: the generator and critic are pretrained separately with conventional losses, then briefly married in a GAN setting for only about 30–60 minutes on 1% to 3% of ImageNet data. This narrow window apparently transfers photorealism without accumulating the usual GAN artifacts, though the author admits he still locates the inflection point by manually inspecting checkpoints for telltale orange skin and “zombie hands.”

Key highlights

• Video colorization works frame-by-frame, yet avoids the flickering objects that plague many GAN-based approaches.
• NoGAN replaces days of progressive GAN training with a tiny, carefully timed joint-training window.
• The author openly warns that historical accuracy remains a huge challenge; convincing colorization does not mean correct colorization, as demonstrated by a mis-colored Golden Gate Bridge example.
• The codebase is archived, with the author pointing to third-party forks, browser ports, and a commercial descendant for continued use.

Caveats

• The repository is explicitly unmaintained; the author has moved on and suggests the project is outdated.
• Automating the NoGAN early-stopping point is still unsolved; the author finds it by hand-checking model saves.
• Colors can be historically inaccurate even when they look plausible.

Verdict

Study it if you are interested in GAN stabilization research or the archaeology of generative image restoration; otherwise, modern diffusion-based tools are likely a better bet for current production work.