How AI Restores 100‑Year‑Old Beijing Footage to 4K: Techniques and Insights

AI‑Powered Restoration of 1920 Beijing Footage

A netizen applied modern AI techniques to a 1920 black‑and‑white archival film originally released by the People’s Daily. The restoration pipeline performed three main operations: colorization, resolution up‑scaling, and frame‑rate interpolation, producing a visually smooth 4K video at 60 fps.

Technical Workflow

The typical workflow for restoring historic film consists of the following stages:

Frame Extraction : The source video is split into individual frames using a tool such as ffmpeg -i input.mp4 -qscale:v 2 frame_%05d.png.

Resolution Upscaling : Each frame is processed with a deep‑learning upscaler (e.g., Topaz Gigapixel AI). The software applies a proprietary convolutional neural network to increase pixel count from 1920×1080 (≈2.07 M pixels) to 3840×2160 (≈8.29 M pixels), a 4× increase (≈6 M additional pixels per frame).

Temporal Interpolation : To raise the frame rate from the original ~15 fps to 60 fps, an AI‑based frame‑interpolation model (e.g., DAIN, RIFE, or the built‑in Gigapixel AI temporal engine) synthesizes intermediate frames by estimating motion vectors and generating plausible in‑between imagery.

Colorization : A neural network trained on large collections of color photographs (e.g., DeOldify or similar) predicts chrominance channels for each grayscale frame, producing plausible color versions while preserving historical tones.

Re‑encoding : The processed frames are recombined into a video file using

ffmpeg -r 60 -i upscaled_%05d.png -c:v libx264 -crf 18 -pix_fmt yuv420p output.mp4

.

Interpolation Methods Overview

Traditional interpolation algorithms are often used as baselines:

Nearest Neighbor : Copies the value of the closest original pixel. Fast but produces blocky, pixelated results.

Bilinear Interpolation : Computes a weighted average of the two nearest pixels, yielding smoother gradients at modest computational cost.

Bicubic Interpolation : Samples 16 neighboring pixels to calculate each new pixel, offering higher fidelity but can introduce slight blur.

Combining bilinear and bicubic interpolation can minimize optical quality loss, but deep‑learning upscalers such as Gigapixel AI further refine the output by learning complex texture patterns.

Colorization Details

Neural‑network colorization models are trained on paired grayscale‑color datasets. During inference, the model predicts the a‑ and b‑channels (in Lab color space) for each input frame, which are then merged with the original luminance channel to reconstruct a full‑color image. This approach preserves historical contrast while adding plausible hues.

References

https://weibo.com/2395607675/J0ZsQnP6a?filter=hot&root_comment_id=4502616270201172&type=comment

https://arstechnica.com/science/2020/02/someone-used-neural-networks-to-upscale-a-famous-1896-video-to-4k-quality/