How AI Restored a Century‑Old Beijing Film in Vibrant Color

AI Revives a Century‑Old Beijing Film

An independent developer known as DaGu Spitzer (real name Hu Wengu), a Beijing‑born game designer, used artificial intelligence to restore a 10‑minute black‑and‑white film from the 1920s that was originally archived by the National Film Board of Canada.

The original footage suffered from low resolution, poor frame rate (6–10 fps), grain, scratches, and a lack of color, making it difficult to view. Spitzer applied a fully AI‑based workflow to address these issues.

Technical Pipeline

First, he increased the frame rate to 60 fps using the open‑source DAIN (Depth‑Aware Video Frame Interpolation) algorithm, originally developed by Shanghai Jiao Tong University researcher Bao Wenbo.

Next, he enhanced the resolution with ESRGAN, a GAN‑based super‑resolution method, bringing the video up to near‑4K quality.

For colorization, he employed DeOldify, which uses a NoGAN training strategy to map grayscale frames to realistic color, producing vivid and historically plausible tones.

Finally, he performed noise reduction with VirtualDub and added a soundtrack of traditional Beijing drum music and other audio sourced from online archives.

Results and Impact

The restored video now runs smoothly at 60 fps, displays in high resolution, and showcases Beijing’s streets, markets, and daily life with authentic‑looking colors. The clip quickly attracted massive attention on Chinese social platforms, garnering over 16 million views, 575 000 likes, and thousands of comments reflecting nostalgia and curiosity.

Spitzer notes that the AI model was trained primarily on foreign historic films, so some color choices are based on intuition rather than precise historical data. He plans to fine‑tune the model with Chinese archival material to improve cultural accuracy.

Broader Applications

The same AI techniques have been used to restore other historic footage, such as the 1895 film “Train Arriving at Station,” which was upscaled to 4K and interpolated to 60 fps using Gigapixel AI, DAIN, and related tools.

These examples demonstrate how deep‑learning‑based restoration can revive lost visual heritage, though challenges like color fidelity and artifact removal remain.