Research map / Video / Video generation
One-Step Diffusion for Detail-Rich and Temporally Consistent 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 DLoRAL, a one-step SD-based real-world video super-resolution method that decouples temporal consistency (C-LoRA with a Cross-Frame Retrieval module) and detail enhancement (D-LoRA) into two LoRA branches trained alternately.
Problem
Real-world video super-resolution with pre-trained Stable Diffusion must balance fine spatial detail against temporal consistency; joint optimization in existing SD-based Real-VSR methods yields suboptimal trade-offs and multi-step inference is slow.
Approach
Decouple the objectives into two LoRA branches in a shared SD UNet: a Consistency-LoRA learns degradation-robust temporal representations via a Cross-Frame Retrieval module (top-k, learnable-threshold attention over warped previous-frame latents), then a frozen-consistency Detail-LoRA is trained with CSD loss; the two stages alternate, and both LoRAs are merged for single-step inference.
Key results
- On VideoLQ, DLoRAL reaches MUSIQ 63.846, CLIPIQA 0.567, MANIQA 0.344 and DOVER 38.505, the best among compared methods on these metrics.
- Inference takes 346 s for 50 512x512 frames on an A100, roughly 10x faster than Upscale-A-Video (3640 s), MGLD (4146 s) and STAR (2830 s) on the 15-step/30-step baselines.
- In a 10-person user study on 12 VideoLQ clips, DLoRAL received 93 votes versus 14 (MGLD), 8 (STAR) and 5 (Upscale-A-Video).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DLoRAL instead of competing Real-VSR methods | Real-VSR · UDM10 | LPIPS: 0.335 -> 0.327 | Helped |
| Use DLoRAL instead of competing Real-VSR methods | Real-VSR · UDM10 | E*warp: 3.510 -> 4.720 | Hurt |
| Use DLoRAL instead of competing Real-VSR methods | Real-VSR · SPMCS | LPIPS: 0.384 -> 0.375 | Helped |
| Use DLoRAL instead of competing Real-VSR methods | Real-VSR · SPMCS | E*warp: 4.080 -> 6.250 | Hurt |
| Use DLoRAL instead of competing Real-VSR methods | Real-VSR · RealVSR | E*warp: 24.600 -> 17.300 | Helped |
| Use DLoRAL instead of competing Real-VSR methods | Real-VSR · VideoLQ | DOVER: 37.580 -> 38.505 | Helped |
| Replace multi-step diffusion VSR with one-step DLoRAL inference | DLoRAL · 50 512×512 frames, ×4 VSR | inference time (s/50 frames): 3640 -> 346; 4146 -> 346; 2830 -> 346 | Helped |
Surprising findings
- Its warping error E*warp is not the best on several datasets (e.g., 6.250 on SPMCS vs 4.080 for STAR), and PSNR/SSIM are lower than RealBasicVSR and STAR on synthetic sets.
Limitations
- The 8x downsampling SD VAE struggles to restore very fine details such as small text.
- The heavy VAE compression may disrupt temporal coherence and make consistency priors harder to extract.
- Warping error does not correlate well with human perception, and DLoRAL's E*warp is higher than some baselines.
Experimental setup
Datasets: REDS Pexels videos LSDIR UDM10 SPMCS RealVSR VideoLQ
Benchmarks: UDM10 SPMCS RealVSR VideoLQ
Metrics: PSNR SSIM LPIPS DISTS BRISQUE MUSIQ CLIPIQA MANIQA DOVER E*warp
Baselines: RealESRGAN StableSR OSEDiff RealBasicVSR Upscale-A-Video MGLD-VSR STAR
Builds on: PiSA-SR OSEDiff Stable Diffusion RealESRGAN classifier score distillation
Closest papers by meaning
- Upscale-A-Video: Temporal-Consistent Diffusion Model for Real-World Video Super-Resolution (CVPR 2024 · similarity 0.90)
- Learning Spatial Adaptation and Temporal Coherence in Diffusion Models for Video Super-Resolution (CVPR 2024 · similarity 0.89)
- DOVE: Efficient One-Step Diffusion Model for Real-World Video Super-Resolution (NeurIPS 2025 · similarity 0.89)
- PatchVSR: Breaking Video Diffusion Resolution Limits with Patch-wise Video Super-Resolution (CVPR 2025 · similarity 0.89)
- SeeClear: Semantic Distillation Enhances Pixel Condensation for Video Super-Resolution (NeurIPS 2024 · similarity 0.89)
- Pixel-level and Semantic-level Adjustable Super-resolution: A Dual-LoRA Approach (CVPR 2025 · 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.