TexTile: A Differentiable Metric for Texture Tileability

Pretrained weights for TexTile [CVPR 2024], a differentiable metric that measures how well a texture image can be tiled with itself without visible seams or repeating artifacts.

Carlos Rodriguez-Pardo, Dan Casas, Elena Garces, Jorge Lopez-Moreno

• Paper: arXiv:2403.12961
• Code: github.com/crp94/textile

Files

File	Description
model.safetensors	Weights in safetensors format. Loaded by the textile-metric package.
textile_v3.pth	Original PyTorch state dict (identical weights).

Usage

pip install textile-metric

import textile
from textile.utils.image_utils import read_and_process_image

loss_textile = textile.Textile()  # downloads model.safetensors from this repo
image = read_and_process_image(YOUR_PATH)
textile_value = loss_textile(image)

The architecture is a ConvNeXt-Base with added linear self-attention layers; see textile/utils/create_model.py in the code repository.

Citation

@inproceedings{Rodriguez-Pardo_2024_CVPR,
    author = {Rodriguez-Pardo, Carlos and Casas, Dan and Garces, Elena and Lopez-Moreno, Jorge},
    title = {TexTile: A Differentiable Metric for Texture Tileability},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    year = {2024}
}