Towards Facial Image Compression with Consistency Preserving Diffusion Prior (2505.05870v1)

Abstract: With the widespread application of facial image data across various domains, the efficient storage and transmission of facial images has garnered significant attention. However, the existing learned face image compression methods often produce unsatisfactory reconstructed image quality at low bit rates. Simply adapting diffusion-based compression methods to facial compression tasks results in reconstructed images that perform poorly in downstream applications due to insufficient preservation of high-frequency information. To further explore the diffusion prior in facial image compression, we propose Facial Image Compression with a Stable Diffusion Prior (FaSDiff), a method that preserves consistency through frequency enhancement. FaSDiff employs a high-frequency-sensitive compressor in an end-to-end framework to capture fine image details and produce robust visual prompts. Additionally, we introduce a hybrid low-frequency enhancement module that disentangles low-frequency facial semantics and stably modulates the diffusion prior alongside visual prompts. The proposed modules allow FaSDiff to leverage diffusion priors for superior human visual perception while minimizing performance loss in machine vision due to semantic inconsistency. Extensive experiments show that FaSDiff outperforms state-of-the-art methods in balancing human visual quality and machine vision accuracy. The code will be released after the paper is accepted.

Towards Facial Image Compression with Consistency Preserving Diffusion Prior

The paper "Towards Facial Image Compression with Consistency Preserving Diffusion Prior" presents an innovative approach to facial image compression by leveraging diffusion models, particularly focusing on preserving high-frequency components that are critical for both perceptual quality and functional performance in downstream tasks. The proposed method, FaSDiff, introduces a stable diffusion prior integrated with novel techniques aimed at enhancing frequency retention and ensuring semantic consistency in the reconstruction process.

Methodological Framework

The authors propose FaSDiff, a facial image compression framework that combines a high-frequency-sensitive compressor and a hybrid low-frequency enhancement module to effectively encode and reconstruct facial images. The core components of FaSDiff include:

• High-Frequency Sensitive Compressor: This component captures intricate details of facial images by preserving high-frequency signals. It employs facial consistency guidance through a variance-weighted loss to uphold semantic consistency, as facial details are paramount for human and machine vision.
• Time-aware High-Frequency Augmentation (TaHFA): Designed to augment the high-frequency control signals in synchronization with the denoising process, TaHFA adapts to later stages of diffusion, enhancing the realism and fidelity of generated images without introducing artifacts that could compromise visual quality.
• Hybrid Low-Frequency Enhancement: Recognizing the limitations of CLIP-based alignment, FaSDiff incorporates facial semantic embeddings to stabilize low-frequency components, ensuring that color and style are consistently reproduced during image generation.

Numerical Results and Performance Evaluation

Extensive experiments show that FaSDiff excels in both perceptual reconstruction quality and performance in facial-related tasks. In comparison to baseline methods, FaSDiff achieves a superior balance between extremely low bitrates and high-quality image reconstruction. It consistently outperforms state-of-the-art techniques, such as HiFiC and Perco, in terms of perceptual metrics (e.g., LPIPS, CLIP-IQA) and machine vision metrics (e.g., FWIoU and gender classification accuracy).

Implications and Future Directions

The implications of this research are twofold. Practically, it offers a method capable of significantly reducing storage requirements for facial image data without compromising image quality, which is crucial for applications like identity verification and social media. Theoretically, FaSDiff advances the understanding of frequency-domain processing in diffusion models, paving the way for more robust generative frameworks in image compression tasks.

Future research could explore further integration of diffusion models with other modalities, extending the framework to video compression tasks or multimodal data synthesis. Additionally, addressing computational efficiency in training and inference stages could enhance the practical deployment of such methods in real-time applications. Integrating AI-driven adaptation to varying image characteristics could also present avenues for dynamic compression strategies, optimizing both bitrate and image fidelity across diverse image datasets.