Traditionally, diffusion models have primarily been employed in image generation for progressively denoising random noise images until a clear and coherent image emerges. However, we have taken a groundbreaking approach by repurposing diffusion models to tackle a different challenge: colorizing black and white images. This innovative utilization of the diffusion model paradigm has resulted in the development of state-of-the-art image colorization models that surpass existing methods in both qualitative and quantitative evaluations.

To train our diffusion-based colorization model, we introduce a conceptual shift by replacing the noise level with the color level. We teach a denoising Unet model to incorporate color incrementally at each iteration step instead of focusing solely on noise reduction. Remarkably, this adaptation for the colorization task has proven to be remarkably successful, yielding a highly robust and powerful colorization model.

By leveraging the inherent capabilities of diffusion models in a novel context, we have unlocked the potential to breathe life and vibrancy into black and white images. Our approach surpasses conventional methods, achieving unparalleled results in terms of colorization quality and fidelity.