Spaces:

---

senmaonk

/

InterLCM

like 2

Running on Zero

This project focuses on blind face restoration (BFR) by leveraging latent consistency models (LCMs) as efficient and semantically consistent diffusion priors. Existing diffusion-model-based BFR methods typically fine-tune diffusion models on restoration datasets to recover low-quality facial images. However, directly applying diffusion models to BFR suffers from limited semantic consistency in identity, structure, and color, as well as inefficient inference caused by hundreds of denoising iterations. These limitations make it difficult to optimize restoration models and hinder the effective use of perceptual losses, which are essential for faithful face restoration.

To address these challenges, this project introduces InterLCM, a restoration framework that exploits the consistency noise-to-data mapping learned by LCMs along the ODE trajectory. By treating low-quality images as intermediate states within the LCM process, InterLCM starts restoration from earlier LCM steps to achieve a better balance between fidelity and perceptual quality. The efficiency of LCM also enables the integration of perceptual loss during training, improving restoration performance, especially in real-world degraded scenarios.

In addition, the project incorporates a Visual Module to extract high-level visual features and a Spatial Encoder to capture fine-grained spatial details. These components help reduce structural and semantic uncertainties, thereby improving the fidelity and realism of restored facial images. Experimental results on both synthetic and real-world datasets show that InterLCM achieves superior restoration quality compared with existing methods while also providing faster inference speed.