Research map / Video / Video segmentation and tracking
TAQ: Static-Deployable Temporal-Aware Quantization for Real-World Video Super-Resolution
Jinwoo Chung, Sangho An, Sungyeop Jung, Jangho Kim
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add sequence-wise calibration with bound ensembling (ablation) | TAQ on RealViformer · REDS (4-bit) | PSNR: 23.1268 -> 23.8104 (PSNR rises by 0.68 dB) | Improved |
| add sequence-wise calibration with bound ensembling (ablation) | TAQ on RealViformer · REDS (4-bit) | TLPIPS: 15.3454 -> 11.3696 | Improved |
| add sequence-wise calibration with bound ensembling (ablation) | TAQ on RealViformer · REDS (4-bit) | TOF: 55.4030 -> 27.5246 | Improved |
| add temporal refinement with temporal consistency loss (ablation) | TAQ on RealViformer · REDS (4-bit) | TLPIPS: 11.3696 -> 10.0249 | Improved |
| add temporal refinement with temporal consistency loss (ablation) | TAQ on RealViformer · REDS (4-bit) | PSNR: 23.8104 -> 23.8949 | Improved |
| add 99.9% percentile clipping (ablation) | TAQ on RealViformer · REDS (4-bit) | PSNR: 22.6066 -> 23.8949 | Improved |
| add 99.9% percentile clipping (ablation) | 2DQuant on RealViformer · REDS (4-bit) | TLPIPS: 17.8592 -> 12.9041 | Improved |
| use arithmetic mean instead of median for bound ensembling (ablation) | TAQ on RealViformer · REDS (4-bit) | PSNR: 23.0800 -> 23.8949 | Improved |
| use arithmetic mean instead of trimmed mean for bound ensembling (ablation) | TAQ on RealViformer · REDS (4-bit) | PSNR: 23.8851 -> 23.8949 | No clear effect |
| use learned aggregation weights instead of uniform mean (ablation) | TAQ on RealViformer · REDS (4-bit) | TLPIPS: 10.0249 -> 10.2315 (PSNR 23.8949 -> 23.8996) | Hurt |
| use TAQ instead of 2DQuant with only 5 calibration sequences | TAQ vs 2DQuant · REDS (4/4-bit, 5 calib seqs) | TLPIPS: 14.09 -> 10.20 | Improved |
| use TAQ instead of 2DQuant on RealBasicVSR backbone | TAQ vs 2DQuant · REDS (4-bit, RealBasicVSR) | TLPIPS: 5.4051 -> 4.2680 | Improved |
Closest papers by meaning
- QBasicVSR: Temporal Awareness Adaptation Quantization for Video Super-Resolution (NeurIPS 2025 · similarity 0.90)
- QuantVSR: Low-Bit Post-Training Quantization for Real-World Video Super-Resolution (AAAI 2026 · similarity 0.89)
- Outlier-Aware Post-Training Quantization for Image Super-Resolution (ICCV 2025 · similarity 0.88)
- PMQ-VE: Progressive Multi-Frame Quantization for Video Enhancement (NeurIPS 2025 · similarity 0.86)
- 2DQuant: Low-bit Post-Training Quantization for Image Super-Resolution (NeurIPS 2024 · similarity 0.86)
- DSAQuant: Denoising-Stage-Aligned Quantization-Aware Training for Video Generation (NeurIPS 2026 · similarity 0.85)
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.
