Controllable Texture Tiling with Transformed RoPE-Enhanced Diffusion Models

Traditional texture mapping in computer graphics often requires labor-intensive 3D UV parameterization. While generative AI models have improved image editing, they struggle with the specific requirements of texture tiling—precisely repeating a reference pattern according to user-defined frequency, orientation, and scale. This paper addresses the challenge of performing high-fidelity, controllable texture tiling within a Diffusion Transformer (DiT) architecture without resorting to destructive pixel-level resampling.

The authors propose a framework that decouples spatial manipulation from content generation using two primary technical innovations:

1. Coordinate-Transformed Rotary Embedding (RoPE): Instead of warping the input texture pixels, the model applies 2D affine transformations directly to the relative positional embeddings of the reference texture. This allows the attention mechanism to implicitly align the reference pattern with the target scene geometry based on coordinate math rather than pixel manipulation.

2. Disjoint Attention Mask: To prevent semantic leakage, the model uses a masking strategy that partitions tokens into reference, background, and target subsets. This ensures that the reference texture remains pristine and that the background is not inadvertently altered by the texture generation process.

Additionally, the framework incorporates explicit geometric and lighting priors (depth and irradiance maps) to ensure the synthesized texture conforms to the original scene's shading and surface structure.

Findings and Significance

Extensive experiments demonstrate that this method outperforms state-of-the-art baselines in both control accuracy and texture fidelity. By operating in the latent space and using coordinate-based injection, the model maintains high-frequency details that are typically lost in cross-attention-based methods. The authors also provide a new benchmark dataset of 15,000 scene-texture pairs to standardize evaluation for this task.