DTI: Dynamic Trajectory Initialization for Generative Face Video Super-Resolution

2026-06-28

AI Summary

This work addresses the limitations of existing generative face video super-resolution methods, which rely on fixed sampling trajectories, incur high inference costs, and often sacrifice fidelity for perceptual quality. The authors reformulate the task as an input-driven directional restoration problem and introduce a dynamic trajectory initialization paradigm coupled with a discriminator-guided mechanism. Built upon a pretrained DiT backbone, their approach incorporates an enhancement-injection conditioning scheme and a discriminator-guided module trained with signal-to-noise ratio alignment, enabling efficient deployment with minimal fine-tuning. The proposed method significantly improves fidelity without compromising perceptual quality, achieving state-of-the-art performance across multiple benchmarks. Furthermore, the study validates LPIPS as the most reliable metric for evaluating face video super-resolution quality.

Abstract

As the most perceptually powerful Face Video Super-Resolution (FVSR) method, existing works in Generative FVSR (GFVSR) mainly exploit the generative prior of pretrained diffusion models. However, viewed as full generation, they suffer from fixed sampling and expensive inference costs if without large-scale auxiliary training. Furthermore, an excessive pursuit of generic perceptual metrics often results in low fidelity. To address these issues, we present Dynamic Trajectory Initialization (DTI) paradigm for GFVSR, which reformulates GFVSR as an input-driven directional restoration. With a novel enhancement-and-injection conditioning mechanism for pretrained DiT backbone, fidelity of our model has been significantly improved without compromising perceptual quality. To dynamically set the starting sampling point, we propose a Discriminative Guide (DG) trained via objective Signal-to-Noise Ratio (SNR) alignment. With only minor model adaptation and fine-tuning, our method achieves a SOTA overall performance across diverse metrics and benchmarks. An analysis of relationship between actual comprehensive quality and common metrics is also conducted, which demonstrates the perception-distortion trade-off and that the LPIPS is the most convincing metric in our case.

Problem

• Face Video Super-Resolution
• Generative Prior
• Perceptual Quality
• Fidelity
• Diffusion Models

Innovation

• Dynamic Trajectory Initialization
• Generative Face Video Super-Resolution
• Diffusion Models
• Fidelity-Perception Trade-off
• Discriminative Guide