Controllable Generative Reference for Stereo Image Compression via Reliability-Aware Gating

Abstract

Stereo image compression fundamentally relies on exploiting cross-view redundancies. We observe that the performance of conventional correspondence-driven methodologies degrades severely in out-of-view (OOV) regions or when reference features are heavily quantized, resulting in blurred textures and prediction failure. In this paper, we propose a novel generative stereo image compression framework driven by a controllable generative reference. Instead of relying on finding explicit correspondences from a degraded reference, we leverage the powerful priors of a pre-trained diffusion model to synthesize a high-quality target-view reference. To bridge the gap between stochastic generative processes and strict stereo consistency, we introduce Joint Semantic-Spatial Control for reference generation. This mechanism synergistically steers the synthesis using a robust geometric anchor from the compressed left view and an ultra-compact semantic condition extracted from the target view. Furthermore, to minimize the impact of structural inconsistency during reference generation on the quality of the reconstructed image, we design a Reliability-Aware Gating module. By dynamically evaluating the synthesized prior in the feature space, this module adaptively regulates generative information flow within the entropy model to improve the reconstructed image quality. Extensive experiments on standard benchmarks demonstrate that our framework achieves state-of-the-art rate-distortion performance, delivering remarkable bitrate savings and high visual quality at low bitrates.