Research map / Image generation / Super-resolution
DiSR-NeRF: Diffusion-Guided View-Consistent Super-Resolution NeRF
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 DiSR-NeRF, which super-resolves a low-resolution NeRF by alternating diffusion-based 2D upscaling (via a new Renoised Score Distillation objective) with NeRF re-training (Iterative 3D Synchronization), using only LR multi-view images.
Problem
Super-resolution NeRF from only low-resolution multi-view images, where independently upscaled 2D images are inconsistent across views, producing blurred NeRF details.
Approach
Alternates between upscaling rendered LR views with a pretrained diffusion upscaler and retraining the NeRF on those upscaled views (I3DS), and uses Renoised Score Distillation, which optimizes latent residuals toward the ancestral sampling trajectory to combine SDS sharpness-free LR consistency with ancestral detail.
Key results
- On NeRF-Synthetic, DiSR-NeRF achieves NIQE 5.386 and LPIPS 0.144, versus 5.847/0.173 for the best baseline IN2N.
- On LLFF, DiSR-NeRF achieves NIQE 5.544 and LPIPS 0.141, versus 6.473/0.157 for IN2N.
- Replacing I3DS with DreamFusion-style optimization raises NIQE to 8.212 on NeRF-Synthetic, showing I3DS is essential.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use RSD instead of SDS for upscaling (ablation) | DiSR-NeRF · NeRF-Synthetic and LLFF | NIQE: 6.273 -> 5.386; 5.903 -> 5.544 | Helped |
| Use RSD instead of SDS for upscaling (ablation) | DiSR-NeRF · NeRF-Synthetic and LLFF | LPIPS (warped): 0.175 -> 0.144; 0.164 -> 0.141 | Helped |
| Use RSD instead of ancestral sampling for upscaling (ablation) | DiSR-NeRF · NeRF-Synthetic and LLFF | NIQE: 5.942 -> 5.386; 6.325 -> 5.544 | Helped |
| Use RSD instead of ancestral sampling for upscaling (ablation) | DiSR-NeRF · NeRF-Synthetic and LLFF | LPIPS (warped): 0.189 -> 0.144; 0.159 -> 0.141 | Helped |
| Use I3DS instead of DreamFusion synchronization (ablation) | DiSR-NeRF · NeRF-Synthetic and LLFF | NIQE: 8.212 -> 5.386; 8.299 -> 5.544 | Helped |
| Use I3DS instead of DreamFusion synchronization (ablation) | DiSR-NeRF · NeRF-Synthetic and LLFF | LPIPS (warped): 0.254 -> 0.144; 0.236 -> 0.141 | Helped |
Surprising findings
- Default ancestral sampling of the diffusion upscaler produces details structurally inconsistent with the LR conditioning, aggravating cross-view inconsistency.
- Directly applying SDS to NeRF renders over small patches yields blurred details and fails to converge.
Limitations
- No quantitative comparison with Super-NeRF because its code is not public; only qualitative comparison.
- No comparison against HR ground truth, since SR details are generative and multiple valid outputs exist.
Experimental setup
Datasets: NeRF-synthetic LLFF
Benchmarks: NeRF-synthetic LLFF
Metrics: NIQE warped LPIPS
Baselines: NGP SD×4 NeRF-SR DreamFusion IN2N Super-NeRF
Builds on: Instant-NGP Score Distillation Sampling Stable Diffusion x4 Upscaler Instruct-NeRF2NeRF NeRF-SR
Closest papers by meaning
- NeRFDeformer: NeRF Transformation from a Single View via 3D Scene Flows (CVPR 2024 · similarity 0.87)
- NerfBaselines: Consistent and Reproducible Evaluation of Novel View Synthesis Methods (NeurIPS 2025 · similarity 0.87)
- SinSR: Diffusion-Based Image Super-Resolution in a Single Step (CVPR 2024 · similarity 0.87)
- Unsupervised Diffusion-Based Degradation Modeling for Real-World Super-Resolution (AAAI 2025 · similarity 0.87)
- When Latents Forget Pixels: Restoring Fidelity in Diffusion Transformer Super-Resolution (NeurIPS 2026 · similarity 0.86)
- MVIP-NeRF: Multi-view 3D Inpainting on NeRF Scenes via Diffusion Prior (CVPR 2024 · similarity 0.86)
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.