How video super-resolution works
Video upscaling isn’t just image upscaling, 30 times a second. Neighbouring frames carry extra information, and keeping detail stable over time is half the job.
Single-frame vs multi-frame
A single-frame approach runs an image super-resolution model on every frame independently. It’s simple, but generated texture differs slightly per frame, which shows up as shimmer. A multi-frame (video) model aligns several neighbouring frames, often with optical flow or deformable convolutions, and fuses them. Small camera and subject motion means each frame samples the scene at slightly different positions, so there is genuinely more information to recover.
Model families
| Model | Year | Idea |
|---|---|---|
| EDVR | 2019 | Deformable alignment and fusion of neighbouring frames |
| BasicVSR / BasicVSR++ | 2021 / 2022 | Bidirectional recurrent propagation; strong, compact baselines |
| RealBasicVSR | 2022 | Trained for real-world degradations, with a pre-cleaning stage |
| Real-ESRGAN (per frame) | 2021 | Image model applied frame by frame; fast, with an anime video variant |
| SeedVR / SeedVR2 | 2025 | Diffusion transformer for video restoration; SeedVR2 runs in one step |
| FlashVSR | 2025 | Diffusion-based, designed for fast streaming super-resolution |
Years refer to the first paper or preprint.
Temporal consistency
The hardest part is staying stable over time. Watch upscaled footage at normal speed, not only as stills: flicker on skin, foliage “boiling” and textures that crawl across surfaces are the tell-tale failures, especially for generative models.
Real-time vs offline
Players and GPUs can upscale during playback (NVIDIA RTX Video Super Resolution, Anime4K shaders). That improves what you see, but creates no file. For a deliverable, you need offline upscaling that writes a new video.
Several of these models can be compared side by side on Upsampler’s video upscaler model comparison.
FAQ
Why not just upscale each frame with an image upscaler?
You can, but each frame is processed independently. Generated detail changes slightly from frame to frame and shows up as flicker or “boiling” texture. Video models look at neighbouring frames to keep detail stable and to recover more real detail.
What is NVIDIA RTX Video Super Resolution?
A real-time upscaling feature on RTX graphics cards that improves streamed video in supported browsers and players. It changes playback only; it doesn’t create an upscaled file.
Are diffusion video upscalers better?
They can produce much richer detail, especially on low-quality sources, but they cost more compute and can invent detail. One-step models such as SeedVR2, and streaming-oriented designs such as FlashVSR, are narrowing the speed gap.