Research map / Video / Video generation
InfVSR: Toward Consistency-Driven Streaming Generative Video Super-Resolution
Ziqing Zhang, Kai Liu, Zheng Chen, Xi Li, Yucong Chen, Bingnan Duan, Linghe Kong, Yulun Zhang
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 |
|---|---|---|---|
| use autoregressive inference with rolling KV-cache instead of simple chunking (ablation) | InfVSR · UDM10 | LPIPS: 0.3178 -> 0.2972 | Improved |
| add overlap and blending aggregation to chunking instead of AR (ablation) | chunked VSR · UDM10 | E*warp / semantic consistency: E*warp 2.20 -> 1.96; (BC+SC)/2 0.9456 -> 0.9456 | Mixed |
| add joint visual guidance (vs none) (ablation) | InfVSR · UDM10 | DISTS: 0.1518 -> 0.1422 | Improved |
| share one guidance across chunks (joint) instead of separate per chunk (ablation) | InfVSR · UDM10 | (BC+SC)/2: 0.9547 -> 0.9578 | Improved |
| use chunk/cache (3,3) instead of (1,1) (ablation) | InfVSR · UDM10 | CLIP-IQA: 0.4755 -> 0.5142 | Improved |
| increase chunk/cache from (3,3) to (5,5) (ablation) | InfVSR · UDM10 | PSNR: 24.86 -> 24.90 | No clear effect |
| keep full KV-cache (infinity,3) instead of rolling length 3 (ablation) | InfVSR · UDM10 | CLIP-IQA: 0.5142 -> 0.5084 | Hurt |
| add temporal smoothness loss Ltemp (ablation) | InfVSR · UDM10 | E*warp: 2.23 -> 1.95 | Improved |
| add patch-wise pixel supervision (ablation) | InfVSR · UDM10 | CLIP-IQA: 0.3997 -> 0.5142 | Improved |
| add Stage-I high-resolution pretraining (ablation) | InfVSR · UDM10 | LPIPS: 0.3125 -> 0.2972 | Improved |
| add cross-chunk DMD loss (ablation) | InfVSR · UDM10 | CLIP-IQA / DOVER: CLIP-IQA 0.5028 -> 0.5142; DOVER 0.7603 -> 0.7826; PSNR 25.04 -> 24.86 | Mixed |
| replace Ltemp with optical-flow alignment loss (ablation) | InfVSR · UDM10 | E*warp: 1.95 -> 2.24 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- SeedVR2 — SeedVR: Seeding Infinity in Diffusion Transformer Towards Generic Video Restoration (CVPR 2025) · better on 4 of 7 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Stream-DiffVSR: Low-Latency Streamable Video Super-Resolution via Auto-Regressive Diffusion (ECCV 2026 · similarity 0.92)
- PatchVSR: Breaking Video Diffusion Resolution Limits with Patch-wise Video Super-Resolution (CVPR 2025 · similarity 0.90)
- One-Step Diffusion for Detail-Rich and Temporally Consistent Video Super-Resolution (NeurIPS 2025 · similarity 0.90)
- SeeClear: Semantic Distillation Enhances Pixel Condensation for Video Super-Resolution (NeurIPS 2024 · similarity 0.89)
- VideoGigaGAN: Towards Detail-rich Video Super-Resolution (CVPR 2025 · similarity 0.89)
- Upscale-A-Video: Temporal-Consistent Diffusion Model for Real-World Video Super-Resolution (CVPR 2024 · similarity 0.89)
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.
