TD-BFR: Truncated Diffusion Model for Efficient Blind Face Restoration

TL;DR

The paper tackles slow sampling and detail loss in diffusion-based blind face restoration by introducing TD-BFR, a three-stage truncated-diffusion framework that progressively restores degraded faces from low-resolution inputs. It combines a Low-resolution Startup (LRS), an Adaptive Degradation Remover (ADR), and a Generative Detail Boost (GDB) to integrate degraded inputs, remove unknown degradations, and exploit pre-trained diffusion priors for high-frequency detail. By using truncated sampling guided by SNR-based breakpoints to connect resolutions, TD-BFR achieves an average speedup of $4.75\times$ while delivering competitive or superior restoration quality on synthetic and real-world datasets. This approach provides a scalable, efficient pathway for high-quality BFR in practical applications.

Abstract

Diffusion-based methodologies have shown significant potential in blind face restoration (BFR), leveraging their robust generative capabilities. However, they are often criticized for two significant problems: 1) slow training and inference speed, and 2) inadequate recovery of fine-grained facial details. To address these problems, we propose a novel Truncated Diffusion model for efficient Blind Face Restoration (TD-BFR), a three-stage paradigm tailored for the progressive resolution of degraded images. Specifically, TD-BFR utilizes an innovative truncated sampling method, starting from low-quality (LQ) images at low resolution to enhance sampling speed, and then introduces an adaptive degradation removal module to handle unknown degradations and connect the generation processes across different resolutions. Additionally, we further adapt the priors of pre-trained diffusion models to recover rich facial details. Our method efficiently restores high-quality images in a coarse-to-fine manner and experimental results demonstrate that TD-BFR is, on average, \textbf{4.75$\times$} faster than current state-of-the-art diffusion-based BFR methods while maintaining competitive quality.

Figures (6)

• Figure 1: (a) presents the mean and variance of $(q(x_t|x_0) - q(x_0))$, and the visualized result at 150 steps for the 64-resolution diffusion model. (b) and (c) shows the corresponding data at 350 and 450 steps for the 256 resolution.
• Figure 2: The Signal-to-noise ratio (SNR) curves of diffusion models at different resolutions. The breakpoints for truncated diffusion models can be selected based on similar SNR values, such as $SNR(t^{64}_{end}) \approx SNR(t^{256}_{begin})$.
• Figure 3: Overall TD-BFR: A three-stage paradigm tailoring for progressive resolution of the input image. ADR serves as a bridge for adaptive handling of unknown degradation and resolution variance, enabling low-resolution integration via LRS and detail generation via GDB.
• Figure 4: Qualitative comparisons on synthetic CelebA datasets. Our method (TD-BFR) achieves higher restoration quality, encompassing more precise facial details with fewer semantic alterations. (Zoom in for the best view)
• Figure 5: Qualitative comparison with SOTA BFR methods on Celeb-Child (first row), LFW-Test (second row), and WIDER-Test (third row).