DeOldify vs DDColor vs BigColor: Measuring Colorization Error with CIEDE2000 (and the Code)
Short answer: On ten gray-converted Prokudin-Gorskii color photos, DeOldify's median whole-image error was 9.55 ΔE2000 (DDColor 9.32, BigColor 9.04), so on average it is no worse than the newer models. The catch is how it gets there: it under-saturates. Across 34 measured zones its mean chroma was 9.5 against 18.9 in the real photos. It also misses skies in some images, and the hosted version we…
DeOldify, DDColor, and BigColor are three colorization models evaluated on ten gray-converted Prokudin-Gorskii photos. DeOldify's median whole-image error was 9.55 ΔE2000, comparable to the newer models, but it tends to under-saturate, with a mean chroma of 9.5 compared to 18.9 in the original images. This model also fails to accurately colorize skies in some images and adds a small icon in the bottom-left corner when using the hosted version.
DeOldify uses a U-Net generator trained with a GAN in a staged recipe called NoGAN: it first pre-trains the generator and a critic to distinguish real from colorized, then runs a short adversarial training. The default artistic model uses a ResNet34 backbone, while the stable model uses ResNet101. The render_factor setting trades rendering resolution for stability. The repository was archived in October 2024.
Colorization error is measured by converting the gray image and the model's output to CIELAB color space, computing the CIEDE2000 color difference between them, and isolating color-only error by forcing the output's lightness to match the original. The key metrics are calculated per pixel in CIELAB, with the median value representing the whole image and hand-boxed zones (sky, vegetation, water, building, painted roof, glazed tile).
The wrong flag identifies cases where the median color differs by 15+ ΔE2000, hue is off by over 40°, or the colorized image appears nearly gray.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.