Research map / Image generation / Autoregressive and latent generation
TexTile: A Differentiable Metric for Texture Tileability
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes TexTile, a differentiable no-reference binary classifier metric for texture tileability built on ConvNext with linear self-attention and tileability-specific augmentation, usable as a loss for texture synthesis.
Problem
Existing texture synthesis evaluation metrics (SSIM, LPIPS, BRISQUE, etc.) do not measure whether a texture tiles seamlessly without repeating artifacts, and none is differentiable for tileability.
Approach
Formulate tileability as a binary classification of a 2x2-tiled texture using a ConvNext backbone augmented with residual linear self-attention blocks, trained with custom augmentations that generate tileable and non-tileable variants; the sigmoid output is differentiable and can be added as a loss term to synthesis methods.
Key results
- Final ConvNext-Base-Att model reaches error 0.064, accuracy 0.982, F1 0.983, AUC 0.997 on the held-out test set, outperforming FID-based distances (AUC 0.764).
- Adding TexTile as a loss raises TexTile score of Heitz et al. from 0.431 to 0.781 and of Nikankin et al. (SinFusion) from 0.388 to 0.798 while keeping perceptual metrics roughly stable.
- Existing perceptual metrics show weak mutual correlation (e.g., SSIM and Si-FID correlate -0.04 / weak), so TexTile offers a distinct tileability signal.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add TexTile loss | Heitz et al. · 14-texture dataset | TexTile: 0.431 -> 0.781 | Helped |
| Add TexTile loss | Nikankin et al. · 14-texture dataset | TexTile: 0.388 -> 0.798 | Helped |
| Add ImageNet pretraining (ablation) | TexTile ConvNext-Base-Att · TexTile test set | Error: 0.459 -> 0.064 | Helped |
| Use NAdam (ablation) | TexTile ConvNext-Base-Att · TexTile test set | Error: 0.096 -> 0.064 | Helped |
| Use Look-Ahead (ablation) | TexTile ConvNext-Base-Att · TexTile test set | Error: 0.082 -> 0.064 | Helped |
| Add negative samples (ablation) | TexTile ConvNext-Base-Att · TexTile test set | Error: 0.319 -> 0.064 | Helped |
| Add color augmentation (ablation) | TexTile ConvNext-Base-Att · TexTile test set | Error: 0.119 -> 0.064 | Helped |
| Add rescales (ablation) | TexTile ConvNext-Base-Att · TexTile test set | Error: 0.105 -> 0.064 | Helped |
| Add flips (ablation) | TexTile ConvNext-Base-Att · TexTile test set | Error: 0.087 -> 0.064 | Helped |
| Add distortions (ablation) | TexTile ConvNext-Base-Att · TexTile test set | Error: 0.094 -> 0.064 | Helped |
| Add AUG T→F (ablation) | TexTile ConvNext-Base-Att · TexTile test set | Error: 0.121 -> 0.064 | Helped |
| Add AUG F→F′ (ablation) | TexTile ConvNext-Base-Att · TexTile test set | Error: 0.087 -> 0.064 | Helped |
Surprising findings
- Self-attention-augmented ConvNext outperformed pure ConvNext, ResNet, VGG, and Swin V2 by a clear margin on the validation error.
- Distribution-based metrics such as GS and MSID perform only marginally better than random for tileability classification.
Limitations
- Does not quantify perceptual quality of textures, only tileability.
- Pretrained on ImageNet and fine-tuned on a manually curated dataset, so may inherit biases and underrepresent some texture types.
- Cannot compensate for limitations of the generative backbone; no human perceptual validation performed.
Experimental setup
Datasets: ImageNet SeamlessGAN 14-texture dataset
Benchmarks: 14-texture dataset from SeamlessGAN TexTile held-out test set
Metrics: binary cross-entropy error accuracy F1 AUC SSIM Si-FID LPIPS DISTS PieAPP BRISQUE
Baselines: SSIM Si-FID LPIPS DISTS PieAPP BRISQUE CLIP-IQA FID GS MSID Heitz et al. neural texture synthesis SinFusion (Nikankin et al.)
Builds on: ConvNext Linear Self-Attention NAdam Lookahead SinFusion
Closest papers by meaning
- PICTURE: PhotorealistIC virtual Try-on from UnconstRained dEsigns (CVPR 2024 · similarity 0.85)
- TextureDreamer: Image-Guided Texture Synthesis Through Geometry-Aware Diffusion (CVPR 2024 · similarity 0.83)
- TriTex: Learning Texture from a Single Mesh via Triplane Semantic Features (CVPR 2025 · similarity 0.83)
- MARBLE: Material Recomposition and Blending in CLIP-Space (CVPR 2025 · similarity 0.83)
- Free-Lunch Color-Texture Disentanglement for Stylized Image Generation (NeurIPS 2025 · similarity 0.83)
- AnyDressing: Customizable Multi-Garment Virtual Dressing via Latent Diffusion Models (CVPR 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.