Generative 2D diffusion models possess an incredible capacity for visual synthesis, but they are notoriously "3D-blind." When asked to generate multiple views of the same object, a standard model often produces "identity drift"—the color of a shirt might change, or a character might grow an extra limb when viewed from the back.
To solve this, a growing body of research utilizes a surprisingly simple geometric trick: multi-view tiling. By arranging different perspectives of a 3D scene into a single 2D grid (often called a "superimage" or "reference sheet"), researchers have found that standard 2D models can be "tricked" into maintaining near-perfect 3D consistency.
The Grid Prior: Leveraging Self-Attention
The core intuition behind tiling is that it transforms a 3D consistency problem into a 2D spatial relationship problem. In a standard transformer-based diffusion model, the self-attention mechanism allows every pixel in an image to "attend" to every other pixel.
When multiple viewpoints—such as the front, side, and back of an object—are tiled into a single $2 \times 2$ grid, the model's internal attention layers naturally coordinate features across these tiles. As Ethan Weber, Aleksander Holynski, and their colleagues demonstrate in NeRFiller, this "Grid Prior" forces the model to treat separate views as a single coherent entity. Because the model processes the entire grid simultaneously, a change in the "front" tile is immediately visible to the "side" tile through the attention mechanism, ensuring that textures and structures align across the virtual 3D space.
Why 2D Models Understand Grids
A recurring observation across these papers is that 2D models seem pre-conditioned to understand multi-view grids. Ethan Weber and Saurabh Saxena suggest that because these models are trained on internet-scale data, they have likely encountered millions of image grids—such as product catalogs or photography collages—that show the same object from different angles.
This latent bias allows models to function as a surrogate for 3D structural understanding. In SIGNeRF, Andreas Engelhardt uses a multi-view "reference sheet" to guide edits, showing that the model's ability to process global context within a single frame ensures that generated features align across perspectives without any explicit 3D modeling.
Applications Across the 3D Pipeline
The tiling strategy has proven effective across various 3D tasks, primarily because it bypasses the need for specialized, expensive 3D-aware architectures.
- Fast Reconstruction: Minghua Liu and Chao Xu in One-2-3-45++ use tiling to generate consistent multi-view images from a single input in under a minute. By processing all views in one composite latent representation, they prevent the "multi-face" artifacts common in sequential generation.
- Interactive Editing: ProteusNeRF, developed by Binglun Wang and Niloy Mitra, uses a $2 \times 2$ "3D-aware image context" to allow users to edit NeRFs using standard 2D tools. The tiling ensures that an edit made to one view is synchronized across the others in seconds.
- Relightable Texturing: In FlashTex, Kangle Deng and Jun-Yan Zhu concatenate canonical views into a grid to generate physically-based textures. This method ensures stylistic coherence across the mesh while separating surface color from lighting.
- Scene Inpainting: For completing missing parts of a 3D scan, OccludeNeRF by Jiazhi Li and colleagues tiles rendered viewpoints to "share" information across occluded areas, maintaining visual harmony even when parts of the scene are hidden in most views.
Geometric Priors and Hierarchies
While simple tiling is powerful, recent work has integrated it with more explicit geometric representations. Rishabh Kabra and João Carreira in unPIC use a hierarchical approach, first predicting camera-relative geometry and then using multi-view superimage tiling to "fill in" the appearance. Similarly, TextMesh by Christina Tsalicoglou uses tiled orthographic views to refine textures on a Signed Distance Function (SDF) backbone, ensuring the final mesh is both geometrically smooth and visually consistent.
By treating a collection of views as a single coherent image, multi-view tiling effectively "borrows" the spatial reasoning of 2D models to solve the fundamental challenge of 3D consistency.