Research map / Image generation / Super-resolution
PBR-SR: Mesh PBR Texture Super Resolution from 2D Image Priors
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 PBR-SR, a zero-shot method that super-resolves mesh PBR texture maps (albedo, roughness, metallic, normal) by optimizing them via differentiable rendering against DiffBIR-upscaled pseudo-ground-truth views, with a robust per-view weighting loss and PBR consistency/TV regularizers.
Problem
Low-resolution PBR textures on 3D meshes look poor under close-up and relighting; existing image SR models operate on baked or per-channel images and produce view- and lighting-inconsistent or physically inconsistent PBR maps, and no paired PBR SR training data exists.
Approach
Initialize HR PBR maps (SR albedo, bicubic-upsampled ARM and normal), then iteratively render multi-view images, upscale them with a pretrained image SR prior (DiffBIR) as pseudo-GT, and optimize the shared PBR texture maps with a learnable robust per-pixel weighting loss plus texture-space PBR consistency and TV constraints.
Key results
- PBR-SR achieves the best PSNR on all PBR channels and renderings at 4x SR (albedo 29.731, roughness 31.602, metallic 31.889, normal 29.088, renderings 28.001), beating supervised CAMixerSR-FT (renderings 27.928).
- Removing the PBR consistency loss drops rendering PSNR from 28.001 to 26.572 in the ablation.
- Optimization takes about 30 minutes per mesh for 2K-to-8K on a single A6000 GPU.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use PBR-SR instead of CAMixerSR-FT | PBR texture super-resolution · PBR mesh collection, ×4 SR | Albedo PSNR: 27.800 -> 29.731 | Helped |
| use PBR-SR instead of CAMixerSR-FT | PBR texture super-resolution · PBR mesh collection, ×4 SR | Roughness PSNR: 30.642 -> 31.602 | Helped |
| use PBR-SR instead of HAT | PBR texture super-resolution · PBR mesh collection, ×4 SR | Metallic PSNR: 30.536 -> 31.889 | Helped |
| use PBR-SR instead of Paint-it SR | PBR texture super-resolution · PBR mesh collection, ×4 SR | Normal PSNR: 28.237 -> 29.088 | Helped |
| use PBR-SR instead of CAMixerSR-FT | PBR texture super-resolution · PBR mesh collection, ×4 SR | Renderings PSNR: 27.928 -> 28.001 | Helped |
| add PBR consistency loss (ablation) | PBR-SR · PBR mesh collection, ×4 SR | Renderings PSNR: 26.572 -> 28.001 | Helped |
| add PBR total variation regularization (ablation) | PBR-SR · PBR mesh collection, ×4 SR | Renderings PSNR: 27.593 -> 28.001 | Helped |
| add robust pixel-wise loss (ablation) | PBR-SR · PBR mesh collection, ×4 SR | Renderings PSNR: 27.666 -> 28.001 | Helped |
| use DiffBIR instead of CAMixerSR as the image SR prior (ablation) | PBR-SR · PBR mesh collection, ×4 SR | Renderings PSNR: 27.466 -> 28.001 | Helped |
| use DiffBIR instead of StableSR as the image SR prior (ablation) | PBR-SR · PBR mesh collection, ×4 SR | Metallic PSNR: 32.258 -> 31.889 | Hurt |
| increase pseudo-GT rendering resolution from 128 to 1024 (ablation) | PBR-SR on Table Clock · Table Clock, ×4 SR | Renderings PSNR: 22.485 -> 24.856 | Helped |
Surprising findings
- The optimization-based Paint-it SR baseline scores lowest on rendering PSNR (23.959) despite strong normal-map results.
- StableSR gets the best metallic PSNR (32.258) in the SR-model ablation, beating DiffBIR on that channel.
Limitations
- Struggles with severely degraded LR textures because natural image priors are limited.
- Slow iterative zero-shot optimization limits real-time use.
- Performance on generated PBR textures is bounded by the quality of the input textures.
Experimental setup
Datasets: Poly Haven PBR meshes (19, for CAMixerSR-FT training) 16-mesh PBR evaluation collection (CGTrader, Poly Haven, glTF samples)
Benchmarks: 16-mesh PBR texture collection (4x SR)
Metrics: PSNR on albedo, roughness, metallic, normal maps PSNR on novel-view renderings
Baselines: SwinIR HAT CAMixerSR CAMixerSR-FT OSEDiff StableSR DiffBIR Paint-it SR
Builds on: DiffBIR Paint-it Cook-Torrance microfacet model
Closest papers by meaning
- IntrinsiX: High-Quality PBR Generation using Image Priors (NeurIPS 2025 · similarity 0.87)
- MeshGen: Generating PBR Textured Mesh with Render-Enhanced Auto-Encoder and Generative Data Augmentation (CVPR 2025 · similarity 0.87)
- TexGaussian: Generating High-quality PBR Material via Octree-based 3D Gaussian Splatting (CVPR 2025 · similarity 0.86)
- PBR-NeRF: Inverse Rendering with Physics-Based Neural Fields (CVPR 2025 · similarity 0.86)
- Paint-it: Text-to-Texture Synthesis via Deep Convolutional Texture Map Optimization and Physically-Based Rendering (CVPR 2024 · similarity 0.85)
- Neural Super-Resolution for Real-time Rendering with Radiance Demodulation (CVPR 2024 · similarity 0.84)
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.