Research map / Video / Video segmentation and tracking
VideoGigaGAN: Towards Detail-rich 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.
Adapts the large-scale GAN image upsampler GigaGAN to video super-resolution by adding temporal modules, flow-guided feature propagation, anti-aliasing (BlurPool) downsampling, and a high-frequency shuttle, yielding detailed yet temporally consistent 4x/8x VSR.
Problem
Video super-resolution models that are temporally consistent produce blurry outputs, while generative image upsamplers produce detailed frames that flicker across time; the paper asks how to get both.
Approach
Inflate a pretrained GigaGAN image upsampler into a video model with temporal attention, add BasicVSR++-style flow-guided recurrent feature propagation, replace strided downsampling with anti-aliased BlurPool, and reinject encoder high-frequency residuals into decoder skip connections (HF shuttle) to recover detail without aliasing flicker.
Key results
- On REDS4 (BI), VideoGigaGAN reaches LPIPS 0.1582 vs. 0.1629 for the best regression baseline IA-RT and 0.1786 for BasicVSR++, at the cost of lower PSNR (30.46 vs 32.89 for IA-RT).
- Ablation on REDS4: flow-guided propagation drops LPIPS from 0.2029 to 0.1551; the full model reaches LPIPS 0.1582 with referenced warping error 2.177e-3.
- Runtime of 295 ms per frame at 320x180 to 1280x720 on REDS4, far faster than diffusion VSR baselines (5,168 to 12,212 ms).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add temporal attention (ablation) | GigaGAN video upsampler · REDS4 | LPIPS: 0.2031 -> 0.2029 | No clear effect |
| Add temporal attention (ablation) | GigaGAN video upsampler · REDS4 | Eref_warp (×10−3): 2.497 -> 2.462 | Helped |
| Add flow-guided feature propagation (ablation) | GigaGAN video upsampler with temporal attention · REDS4 | LPIPS: 0.2029 -> 0.1551 | Helped |
| Add flow-guided feature propagation (ablation) | GigaGAN video upsampler with temporal attention · REDS4 | Eref_warp (×10−3): 2.462 -> 2.187 | Helped |
| Add BlurPool anti-aliasing blocks (ablation) | VideoGigaGAN with flow-guided propagation · REDS4 | LPIPS: 0.1551 -> 0.1621 | Hurt |
| Add BlurPool anti-aliasing blocks (ablation) | VideoGigaGAN with flow-guided propagation · REDS4 | Eref_warp (×10−3): 2.187 -> 2.152 | Helped |
| Add high-frequency shuttle (ablation) | VideoGigaGAN with BlurPool · REDS4 | LPIPS: 0.1621 -> 0.1582 | Helped |
| Add high-frequency shuttle (ablation) | VideoGigaGAN with BlurPool · REDS4 | Eref_warp (×10−3): 2.152 -> 2.177 | Hurt |
| Use VideoGigaGAN instead of the strongest LPIPS baseline | Video super-resolution · REDS4, BI | LPIPS: 0.1659 -> 0.1582 | Helped |
| Use VideoGigaGAN instead of the strongest LPIPS baseline | Video super-resolution · Vimeo-90K-T, BI | LPIPS: 0.1428 -> 0.1120 | Helped |
| Use VideoGigaGAN instead of the strongest LPIPS baseline | Video super-resolution · Vid4, BI | LPIPS: 0.2188 -> 0.1925 | Helped |
| Use VideoGigaGAN instead of the strongest LPIPS baseline | Video super-resolution · UDM10, BI | LPIPS: 0.1148 -> 0.1060 | Helped |
Surprising findings
- Adding LPIPS loss to BasicVSR++ and RVRT training causes checkerboard artifacts or divergence, so the GAN-style perceptual loss does not transfer directly to regression VSR.
- Standard warping error favors blurry outputs; bicubic upsampling scores lower than ground truth, motivating the referenced warping error.
Limitations
- PSNR and SSIM are substantially lower than regression-based methods.
- Referenced warping error is still biased toward blurry results, and the paper leaves a better temporal consistency metric to future work.
- Model is larger (369M parameters) than most regression VSR baselines.
Experimental setup
Datasets: REDS Vimeo-90K Vid4 UDM10
Benchmarks: REDS4 Vimeo-90K-T Vid4 UDM10
Metrics: LPIPS PSNR SSIM warping error referenced warping error
Baselines: BasicVSR++ BasicVSR IconVSR TTVSR VRT RVRT EvTexture StableVSR MIA-VSR UAV DiffIR2VR-Zero VEnhancer
Builds on: GigaGAN BasicVSR++ BlurPool StyleGAN2 Imagen Video
Closest papers by meaning
- Arbitrary-Scale Video Super-resolution Guided by Dynamic Context (AAAI 2024 · similarity 0.89)
- Sequence Matters: Harnessing Video Models in 3D Super-Resolution (AAAI 2025 · similarity 0.88)
- One-Step Diffusion for Detail-Rich and Temporally Consistent Video Super-Resolution (NeurIPS 2025 · similarity 0.88)
- SAVSR: Arbitrary-Scale Video Super-Resolution via a Learned Scale-Adaptive Network (AAAI 2024 · similarity 0.88)
- PatchVSR: Breaking Video Diffusion Resolution Limits with Patch-wise Video Super-Resolution (CVPR 2025 · similarity 0.87)
- SeeClear: Semantic Distillation Enhances Pixel Condensation for Video Super-Resolution (NeurIPS 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.