Research map / Video / Video generation
DTG-Restore: Training-Free Diffusion Refinement for Generative Video Super-Resolution
Hidir Yesiltepe, Koutilya PNVR, Gaurav Pathak, Navaneeth Bodla, Bharat Singh, Pinar Yanardag, Jinrong Xie
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 exponential-annealed decoupled time guidance instead of standard CFG (Delta=0) (ablation) | DTG-Restore · GenWarp480 (heavy distortion, schedule ablation) | Quality: 3.91 -> 4.58 | Improved |
| anneal Delta (exponential) instead of constant Delta=0.3 (ablation) | DTG-Restore · GenWarp480 (light distortion) | Quality: 4.03 -> 4.64 | Improved |
| use constant lookahead Delta=0.3 instead of Delta=0 (ablation) | Std. CFG · GenWarp480 (light distortion) | Quality: 4.12 -> 4.03 | Hurt |
| use exponential instead of cosine annealing (ablation) | DTG-Restore · GenWarp480 (heavy distortion, schedule ablation) | Quality: 4.47 -> 4.58 | Improved |
| use cosine instead of linear annealing (ablation) | DTG annealing · GenWarp480 (heavy distortion, schedule ablation) | Quality: 4.41 -> 4.47 | Improved |
| replace SDEdit single-point resampling (t_start=0.5) with DTG per-step decoupling (ablation) | DTG-Restore · GenWarp480 | LAION AP: 4.38 -> 4.64 | Improved |
| replace SDEdit (t_start=0.5) with DTG (ablation) | DTG-Restore · GenWarp480 | Warp score: 0.118 -> 0.071 | Improved |
| deeper SDEdit resampling t_start 0.3 to 0.7 (ablation) | SDEdit · GenWarp480 | Warp score / CLIP-IQA: 0.142 -> 0.097 warp; CLIP-IQA 0.489 -> 0.478 | Mixed |
| DTG-Restore versus strongest reconstruction baseline RealViformer | DTG-Restore · UDM10 | PSNR: 26.78 -> 25.61 | Hurt |
| DTG-Restore versus RealViformer on reference-free MANIQA | DTG-Restore · GenWarp480 | MANIQA: 0.293 -> 0.314 | Improved |
| DTG-Restore versus SeedVR2 on NIQE | DTG-Restore · GenWarp480 | NIQE: 3.915 -> 4.337 | Hurt |
| DTG-Restore versus Venhancer in user study motion smoothness | DTG-Restore · User study | Motion Smoothness: 3.80 -> 4.52 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- SeedVR2-7B — SeedVR: Seeding Infinity in Diffusion Transformer Towards Generic Video Restoration (CVPR 2025) · better on 1 of 3 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
- Upscale-A-Video: Temporal-Consistent Diffusion Model for Real-World Video Super-Resolution (CVPR 2024 · similarity 0.90)
- One-Step Diffusion for Detail-Rich and Temporally Consistent Video Super-Resolution (NeurIPS 2025 · similarity 0.89)
- Inference-Time Text-to-Video Alignment with Diffusion Latent Beam Search (NeurIPS 2025 · similarity 0.88)
- GenFusion: Closing the Loop between Reconstruction and Generation via Videos (CVPR 2025 · similarity 0.88)
- Training-Free Efficient Video Generation via Dynamic Token Carving (NeurIPS 2025 · similarity 0.88)
- DSAQuant: Denoising-Stage-Aligned Quantization-Aware Training for Video Generation (NeurIPS 2026 · similarity 0.88)
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.
