Interactive Video Super-Resolution via Sparse Keyframe Propagation
Most video upscalers are black boxes. You press go and get whatever the model decides, with no way to fix a bad frame. SparkVSR gives you the wheel. Feed it a handful of high-quality “keyframes” (upscaled with any tool you like) and it propagates that exact look across the entire clip — or skip keyframes entirely and let it run fully automatic. Built by researchers at Texas A&M University and YouTube/Google.
🏆 Beats baseline VSR models by up to +24.6% on CLIP-IQA, +21.8% on DOVER, and +5.6% on MUSIQ
📹 See it in action
What it does
SparkVSR is a video super-resolution model that takes low-quality video and reconstructs it at higher resolution with an unusual twist. Most VSR tools are "black boxes": you press go, and you get whatever the model decides to produce, with no way to fix unexpected artifacts. SparkVSR lets you steer the output by providing a handful of already-upscaled keyframes (made with any image super-resolution tool, a commercial API like nano-banana-pro, or an open-source model like PiSA-SR). The model then propagates the look of those keyframes across the whole video, while still respecting the motion and content of the original low-quality footage. If you don't have keyframes, it also works fully blind. Built on top of CogVideoX 1.5 (a 5B image-to-video diffusion model), SparkVSR was developed by researchers at Texas A&M University and YouTube/Google.
Problem it solves
- Lack of control in video upscaling – Traditional VSR tools give you no way to influence what the upscaled video looks like. SparkVSR lets you guide the result with one or more reference keyframes you've already cleaned up the way you want.
- Fixing unexpected artifacts – If a model produces an odd-looking face or texture, you can intervene by upscaling one keyframe yourself (or with another tool) and letting SparkVSR propagate that look forward and backward through the clip.
- Mixing the best of image and video tools – You can use the strongest available image super-resolution tools (which are often sharper than video models) on a few frames, then use SparkVSR to spread that quality to the whole video while keeping motion coherent.
- Works with or without keyframes – Includes a "no-ref" blind mode for fully automatic use, so it can also be used as a drop-in VSR tool without any extra steps.
- Other keyframe-driven video tasks – Beyond upscaling, the same framework also handles tasks like old-film restoration and video style transfer
Input/Output
- Input:
- A low-quality / low-resolution video clip
- (Optional) A single high-quality reference image — typically one frame of the video that you've already cleaned up or upscaled separately. SparkVSR uses it to guide the look of the entire upscaled video
- Output:
- An upscaled high-resolution video at the chosen scale factor
- Smooth, temporally consistent motion that stays grounded in the original low-quality video
- When a reference image is provided: visually consistent with the look/style of that reference
- Parameters:
- Scale factor - choose 2×, 3×, or 4× upscaling
- Reference frame index (Advanced) - the 0-based index of the input video frame that the reference image corresponds to. Ignored when no reference image is provided. Default:
0 - Random seed (Advanced) - set a number for reproducible results, or
-1for random
Accuracy & Speed
Output quality | Beats baseline VSR models by up to +24.6% on CLIP-IQA, +21.8% on DOVER, and +5.6% on MUSIQ perceptual benchmarks |
Upscale factors | 2×, 3×, or 4× |
Hardware needed (training) | 4× A100 GPUs |
Hardware needed (inference) | Single GPU (built on a 5B-parameter diffusion video model — recommended for high-VRAM GPUs) |
Versatility | Out-of-the-box support for old-film restoration and video style transfer |
How SparkVSR compares
SparkVSR is one of four VSR models available in CNAPS Studio’s Video Upscaling lineup — pick based on how much control you need:
SparkVSR | FlashVSR-v1.1 | SeedVR2-3B / 7B | |
Best for | Creative control via reference keyframes | Fastest, near real-time upscaling | High-quality automatic restoration |
Reference-image steering | ✅ Yes (optional) | ❌ No | ❌ No |
Blind / no-reference mode | ✅ Yes | ✅ Yes | ✅ Yes |
Scale factors | 2×, 3×, 4× | up to 4× | up to 4× |
Also handles | Old-film restoration, style transfer | — | — |
All four run inside CNAPS Studio — compare them side-by-side with the Video Compare tool, or chain them into a larger pipeline.
Model Source
- HuggingFace
- License:
Compliance & Provenance
Provider | Open-source |
Provider type | Specialized |
License | |
EU AI Act risk class | Minimal Risk |
Art. 50 transparency | Not applicable |
Region availability | Available globally |
Training data summary | Pending — provider has not yet published per Art. 53(d) |
For more on how we classify models and mark outputs, see our AI Policy.