RAWVAR: Visual Autoregressive Modeling for Color-Faithful Low-Light RAW Image Restoration

Visual autoregressive (VAR) modeling provides a promising generative solution for RGB image restoration, owing to its strong ability to generate realistic details and recover rich scene information. Nevertheless, directly applying VAR to low-light RAW image restoration still presents two main challenges. First, RGB-pretrained VAE tokenizers are mismatched with RAW sensor statistics, causing color bias after domain transfer. Second, VAR requires tiled inference for high-resolution RAW images, where independent token prediction across tiles may introduce color discrepancies that are further amplified by the Image Signal Processing (ISP) pipeline. To address these issues, we propose RAWVAR, the first visual autoregressive framework for low-light RAW image restoration. For effective domain transfer, we design a Cross-Domain Codebook (CDC), which freezes the RGB-pretrained encoder and original codebook while appending a trainable RAW-domain codebook, forming a cross-domain latent space for texture-preserving and color-faithful RAW representation. To mitigate color inconsistency during tiled inference, we further design a Position-aware Token Alignment (PTA) strategy. During coarse-to-fine prediction, PTA aligns the token logits of semantically corresponding regions across tiles, ensuring consistent predictions for shared content. Furthermore, a Global Color Refinement (GCR) module globally refines the discrete latents to enhance image-wise color consistency. Compared with diffusion-based methods, RAWVAR provides a more efficient generative restoration pipeline by avoiding iterative denoising. Extensive experiments show that RAWVAR produces color-faithful and perceptually superior results, outperforming existing generative restoration methods.