ViewMask-1-to-3: Multi-View Consistent Image Generation via Multimodal Discrete Diffusion Models

A research team has introduced ViewMask-1-to-3, a method that applies discrete diffusion models to multi-view image generation, a task long dominated by continuous approaches. The framework represents each viewpoint as discrete visual tokens and generates consistent views through iterative token unmasking, without relying on specialized 3D architectures [1]. The work, submitted to arXiv in December 2025 and last revised in June 2026, comes from Ruishu Zhu and colleagues [1]. ViewMask-1-to-3 formulates multi-view synthesis as a discrete sequence modeling problem, using visual tokens obtained through MAGVIT-v2 tokenization [2]. By unifying language and vision through masked token prediction, the approach enables progressive generation of multiple viewpoints via iterative token unmasking with text input [3]. Simple random masking combined with self-attention naturally encourages cross-view consistency, eliminating the requirement for complex 3D geometric constraints or specialized attention architectures [2]. The method outperformed the baseline on the GSO and 3D-FUTURE benchmarks, ranking first on average across standard image metrics including PSNR, SSIM, and LPIPS [3]. On the 3D-FUTURE dataset, ViewMask-1-to-3 achieved a 10.6% higher Intersection over Union than continuous diffusion models [1]. The proposed framework can also be extended to support text-to-image generation and multimodal understanding, pointing toward a more unified paradigm for multimodal tasks [2]. Multi-view generation from a single image and text description has remained challenging due to the difficulty of maintaining geometric consistency across different viewpoints [3]. Prior work has explored various strategies to address this. The AR-1-to-3 method, for instance, proposed a next-view prediction paradigm that generates views close to the input first, then uses them as contextual information to progressively synthesize farther views [4]. The Consistent-1-to-3 framework decomposed novel view synthesis into two stages: transforming observed regions and hallucinating unseen regions, employing epipolar-guided attention to incorporate geometry constraints [5]. ViewMask-1-to-3 departs from these approaches by operating in a discrete token space rather than a continuous latent space, demonstrating that discrete diffusion provides a viable and simpler alternative to existing multi-view generation methods [3].