Research map / 3D vision / 3D generation
Sequence Matters: Harnessing Video Models in 3D Super-Resolution
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 reordering unordered LR multi-view images into video-like sequences (greedy and adaptive-length subsequence ordering) so pretrained VSR models (PSRT) can upsample them without fine-tuning, then trains 3DGS on the upsampled HR views.
Problem
3D super-resolution from low-resolution multi-view images requires view-consistent upsampling; prior VSR-based approaches need smooth rendered LR trajectories from 3DGS, which cause artifacts and require fine-tuning.
Approach
Instead of rendering smooth trajectories from an LR 3DGS, order the ground-truth LR training images directly into video-like sequences using a greedy nearest-neighbor on ORB feature or pose similarity, split into adaptive-length subsequences with multiple thresholds, and feed them to a pretrained VSR model without fine-tuning.
Key results
- On Blender (x4 to x1), Ours-ALS reaches 31.41 PSNR / 0.9520 SSIM / 0.0540 LPIPS, above SwinIR+3DGS (30.77) and SRGS (30.83 reported) and the best among compared baselines.
- On Mip-NeRF 360 (x8 to x2), Ours-ALS achieves 27.02 PSNR / 0.7741 SSIM / 0.2774 LPIPS, versus SwinIR 26.80 PSNR and SRGS 26.88 PSNR.
- Adaptive-length subsequences improve over the simple greedy ordering on every Blender object average, e.g., Lego PSNR rises from 33.30 to 34.73.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use adaptive-length subsequences instead of simple greedy ordering (ablation) | PSRT-based 3D super-resolution · Blender | PSNR: 31.35 -> 31.41 | Helped |
| use adaptive-length subsequences instead of simple greedy ordering (ablation) | PSRT-based 3D super-resolution · Blender | SSIM: 0.9513 -> 0.9520 | Helped |
| use adaptive-length subsequences instead of simple greedy ordering (ablation) | PSRT-based 3D super-resolution · Blender | LPIPS: 0.0547 -> 0.0540 | Helped |
| use adaptive-length subsequences instead of simple greedy ordering (ablation) | PSRT-based 3D super-resolution · Mip-NeRF 360 | PSNR: 27.02 -> 27.02 | No clear effect |
| use adaptive-length subsequences instead of simple greedy ordering (ablation) | PSRT-based 3D super-resolution · Mip-NeRF 360 | SSIM: 0.7747 -> 0.7741 | Hurt |
| use adaptive-length subsequences instead of simple greedy ordering (ablation) | PSRT-based 3D super-resolution · Mip-NeRF 360 | LPIPS: 0.2790 -> 0.2774 | Helped |
| use adaptive-length subsequences instead of processing images individually (ablation) | VRT-based 3D super-resolution · Blender | PSNR: 31.20 -> 31.37 | Helped |
| use adaptive-length subsequences instead of processing images individually (ablation) | IART-based 3D super-resolution · Blender | PSNR: 31.10 -> 31.35 | Helped |
| use adaptive-length subsequences instead of simple greedy ordering (ablation) | VRT-based 3D super-resolution · Blender | PSNR: 31.25 -> 31.37 | Helped |
| use adaptive-length subsequences instead of simple greedy ordering (ablation) | IART-based 3D super-resolution · Blender | PSNR: 31.32 -> 31.35 | Helped |
| use VSR with reference frames rather than processing images individually (ablation) | PSRT-based 3D super-resolution · Blender | PSNR: 31.10 -> 31.41 | Helped |
| use adaptive-length subsequences rather than the last 25% of simple-greedy sequences (ablation) | PSRT-based 3D super-resolution · Blender | PSNR: 31.32 -> 31.41 | Helped |
Surprising findings
- VSR models perform well even on sequences lacking precise spatial alignment, so exact smooth trajectories are not required.
- Rendering smooth LR videos from a 3DGS trained on LR images (Render-SR) underperforms plain SwinIR upsampling on Blender (28.90 vs 30.77 PSNR).
Limitations
- Misalignments in ORB-based ordering increase with sequence length, degrading quality on later frames of long greedy sequences.
- Blender and Mip-NeRF 360 results rely on hand-chosen thresholds and metric choices; PSNR prioritized over perceptual metrics.
- Several baselines' numbers are copied from original papers because code is unavailable, limiting direct comparison.
Experimental setup
Datasets: NeRF-synthetic Blender Mip-NeRF 360 Vimeo-90K
Benchmarks: NeRF-synthetic Blender Mip-NeRF 360
Metrics: PSNR SSIM LPIPS FVD
Baselines: Bicubic SwinIR Render-SR NeRF-SR ZS-SRT CROP FastSR-NeRF DiSR-NeRF SRGS GaussianSR SuperGaussian
Builds on: 3D Gaussian Splatting PSRT SuperGaussian 3DGS sub-pixel constraint
Closest papers by meaning
- SR3R: Rethinking Super-Resolution 3D Reconstruction With Feed-Forward Gaussian Splatting (CVPR 2026 · similarity 0.91)
- Look Back and Forth: Video Super-Resolution With Explicit Temporal Difference Modeling (CVPR 2022 · similarity 0.89)
- VideoGigaGAN: Towards Detail-rich Video Super-Resolution (CVPR 2025 · similarity 0.88)
- SAVSR: Arbitrary-Scale Video Super-Resolution via a Learned Scale-Adaptive Network (AAAI 2024 · similarity 0.88)
- One-Step Diffusion for Detail-Rich and Temporally Consistent Video Super-Resolution (NeurIPS 2025 · similarity 0.88)
- SinSR: Diffusion-Based Image Super-Resolution in a Single Step (CVPR 2024 · similarity 0.87)
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.