Research map / Image generation / Editing and style transfer
Generating Non-Stationary Textures using Self-Rectification
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 self-rectification, a training-free method that refines a user's lazily edited texture into a coherent non-stationary texture by injecting self-attention KV features from a pre-trained Stable Diffusion during structure-preserving inversion and fine texture sampling.
Problem
Example-based non-stationary texture synthesis that respects user layout edits of a reference exemplar while preserving the reference's local characteristics; prior GAN-based methods need long per-texture optimization and lack controllability.
Approach
Users roughly edit the reference with standard image tools; a two-stage (coarse-to-fine) diffusion procedure inverts the rough target and the reference, then injects self-attention K/V features into the inversion and sampling so the output keeps the target's global structure while matching the reference's local textures, with augmented reference copies concatenated into the attention keys/values.
Key results
- Self-rectification produces outputs that respect user edits while reproducing reference textures, whereas TexExp fails to capture fine details and GCD Loss does not always conform to the target layout (qualitative comparison only).
- Default setting uses Stable Diffusion v1.4 with about three minutes per 512x512 image on a single Quadro P6000 GPU.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use self-rectification instead of TexExp for edited non-stationary texture targets | Non-stationary texture synthesis · Non-stationary textures | fine texture reproduction | Helped |
| Use self-rectification instead of GCD Loss optimization for edited targets | Non-stationary texture synthesis · Non-stationary textures | layout conformity | Helped |
| Remove KV injection during texture sampling (ablation) | Self-Rectification · Non-stationary textures | visual quality | Hurt |
| Increase KV-injection sampling steps by using smaller S values (ablation) | Self-Rectification · Non-stationary textures | similarity to reference texture | Helped |
| Increase S to reserve more of the target layout (ablation) | Self-Rectification · Non-stationary textures | structural artifacts | Hurt |
| Set S1 to 10–20 and S2 to 5 (ablation) | Self-Rectification · Non-stationary textures | balance between layout and texture details | Helped |
| Add KV injection during structure-preserving inversion (ablation) | Self-Rectification · Non-stationary textures | visual quality | Helped |
| Vary P1 and P2 for structure-preserving inversion (ablation) | Self-Rectification · Non-stationary textures | visual quality | No clear effect |
| Add flipped or rotated reference images through data augmentation (ablation) | Self-Rectification · Directional textures | layout conformity | Helped |
Surprising findings
- Data augmentation is often unnecessary because pre-trained Stable Diffusion already synthesizes rotated patterns, except for textures with dominant directional structure.
Limitations
- Inference takes about three minutes per 512x512 image.
- Evaluation is qualitative only, with no quantitative metrics reported in the main text.
- Extension to large-scale textures is left for future work.
Experimental setup
Datasets: non-stationary textures released by Zhou et al. (TexExp)
Benchmarks: non-stationary textures from TexExp (Zhou et al.)
Metrics: qualitative visual comparison
Baselines: TexExp GCD Loss
Builds on: Stable Diffusion DDIM inversion cross-image attention feature injection (Tumanyan et al.) TexExp Plug-and-Play
Closest papers by meaning
- TextureDreamer: Image-Guided Texture Synthesis Through Geometry-Aware Diffusion (CVPR 2024 · similarity 0.86)
- One-Shot Reference-based Structure-Aware Image to Sketch Synthesis (AAAI 2025 · similarity 0.84)
- Single Mesh Diffusion Models with Field Latents for Texture Generation (CVPR 2024 · similarity 0.84)
- GenesisTex: Adapting Image Denoising Diffusion to Texture Space (CVPR 2024 · similarity 0.84)
- Free-Lunch Color-Texture Disentanglement for Stylized Image Generation (NeurIPS 2025 · similarity 0.84)
- Prune and Repaint: Content-Aware Image Retargeting for any Ratio (NeurIPS 2024 · 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.