Research map / Video / Video segmentation and tracking
NegVSR: Augmenting Negatives for Generalized Noise Modeling in Real-world Video 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 NegVSR, a real-world video super-resolution method that mixes unsupervised sequential noise extracted from an OOD video dataset with patch-based random-rotation negative augmentation, trained with an augmented negative guidance loss.
Problem
Real-world video super-resolution struggles with unknown, complex degradation; existing degradation pipelines and independent per-image noise mixing fail to model realistic, temporally consistent video noise, leaving the VSR model vulnerable to out-of-distribution noise.
Approach
Extract temporally consistent noise sequences from OOD video (low-variance windows), mix them into LR training video, and apply patch-based random central rotation as negative augmentation; a consistency loss aligns outputs on negatives with outputs on positives.
Key results
- On VideoLQ, NegVSR achieves NIQE 3.188 and BRISQUE 22.255, better than the original RealBasicVSR (NIQE 3.699, BRISQUE 24.700).
- On the FLIR testing set, NegVSR achieves NIQE 5.225 vs 6.096 for the original RealBasicVSR, with fewer parameters (4.8M vs 6.3M).
- Ablation on VideoLQ: sequential noise (w/ Noise Sequences) reduces NIQE from 3.936 (baseline) to 3.215 with L_Aug-P only; adding NegMix with L_Aug-NP gives 3.188.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use NegVSR instead of the strongest reported baseline | RealBasicVSR, original · VideoLQ | NIQE: 3.699 -> 3.188 | Helped |
| Use NegVSR instead of the strongest reported baseline | RealBasicVSR, original · VideoLQ | BRISQUE: 24.700 -> 22.255 | Helped |
| Use NegVSR instead of the strongest reported baseline | RealBasicVSR, original · VideoLQ | PI: 3.755 -> 3.416 | Helped |
| Use NegVSR instead of the strongest reported baseline | RealBasicVSR, original · VideoLQ | NRQM: 6.313 -> 6.465 | Helped |
| Use NegVSR instead of the strongest reported baseline | RealBasicVSR, original · FLIR testing | NIQE: 6.096 -> 5.225 | Helped |
| Use NegVSR instead of the strongest reported baseline | RealBasicVSR, original · FLIR testing | BRISQUE: 28.428 -> 20.702 | Helped |
| Use NegVSR instead of the strongest reported baseline | RealBasicVSR, original · FLIR testing | PI: 4.765 -> 4.201 | Helped |
| Use NegVSR instead of the strongest reported baseline | RealBasicVSR, original · FLIR testing | NRQM: 6.829 -> 6.973 | Helped |
| Add independently sampled noise (ablation) | RealBasicVSR · VideoLQ | NIQE: 3.936 -> 3.643 | Helped |
| Replace independent noise with sequential real-world noise (ablation) | RealBasicVSR with noise · VideoLQ | NIQE: 3.643 -> 3.215 | Helped |
| Add sequential real-world noise (ablation) | RealBasicVSR · VideoLQ | NIQE: 3.936 -> 3.215 | Helped |
| Apply NegMix to sequential-noise training without augmented negative guidance loss (ablation) | RealBasicVSR with noise sequences · VideoLQ | NIQE: 3.215 -> 3.312 | Hurt |
Surprising findings
- Independent per-frame noise mixup ('Mixup Noise') gave the worst NIQE among the mixing methods compared, worse than the baseline mixing variants.
Limitations
- Inference speed remains a limitation; authors plan lightweight structures for real-time VSR.
- Evaluation relies on no-reference IQA metrics (NIQE, BRISQUE, NRQM, PI) without ground-truth HR references, and uses a subset of frames.
Experimental setup
Datasets: REDS FLIR training set (noise source, OOD video) VideoLQ (test) FLIR testing set
Benchmarks: VideoLQ FLIR testing dataset
Metrics: NIQE BRISQUE PI NRQM Params Runtime
Baselines: Bicubic DAN BSRGAN Real-ESRGAN RealVSR DBVSR RealBasicVSR
Builds on: RealBasicVSR Mixup Real-ESRGAN high-order degradation BasicVSR++
Closest papers by meaning
- Unsupervised Diffusion-Based Degradation Modeling for Real-World Super-Resolution (AAAI 2025 · similarity 0.87)
- VideoGigaGAN: Towards Detail-rich Video Super-Resolution (CVPR 2025 · similarity 0.87)
- DOVE: Efficient One-Step Diffusion Model for Real-World Video Super-Resolution (NeurIPS 2025 · similarity 0.87)
- Sequence Matters: Harnessing Video Models in 3D Super-Resolution (AAAI 2025 · similarity 0.86)
- SAVSR: Arbitrary-Scale Video Super-Resolution via a Learned Scale-Adaptive Network (AAAI 2024 · similarity 0.86)
- SEAL: A Framework for Systematic Evaluation of Real-World Super-Resolution (ICLR 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.