Google open-sources Spark VSSR for video upscaling
Google just open-sourced Spark VSSR, a powerful video super-resolution framework capable of sharpening degraded clips and restoring old films. The release includes inference models, full training code, and datasets, though running the 42.2GB package demands a high-end GPU. Basically, creators get state-of-the-art video upscaling that they can run and fine-tune locally.
Spark VSSR, released on GitHub as SparkVSR, is an open-source video super-resolution framework built to upscale blurry footage into crisp, high-definition video. It handles everything from wildlife close-ups and landscape scenes to aged film restorations and stylized 3D animations, reconstructing fine architectural textures that older tools tend to smudge.
What it can do
Most video upscalers behave like black boxes. When the model introduces weird artifacts or hallucinations, you typically have zero control and simply have to live with the flawed output.
Spark VSSR tackles this by turning sparse keyframes into an interactive steering wheel.
You can take a handful of keyframes, enhance them with any standard image upscaler of your choice, and let Spark VSSR propagate those high-resolution details across the entire video timeline while preserving the original motion.
On standard benchmarks, it beats baseline models by up to 24.6% on CLIP-IQA, 21.8% on DOVER, and 5.6% on MUSIQ. Compared to earlier tools like Star, CDVR, or Flash VSSR, it delivers noticeably better perceptual sharpness and frame-to-frame consistency.
Hardware and specs
The release is unusually complete for an academic project. Google and Texas A&M researchers shared the inference weights, complete training code, dataset scripts, and an official ComfyUI extension.
The catch is the hardware footprint. The full release package weighs around 42.2 GB, which means running it locally requires a high-end GPU with plenty of VRAM.
On the software side, you need Python 3.10+, PyTorch 2.5.0+, and the standard Diffusers library. Everything is hosted on GitHub under taco-group/SparkVSR, with community deployments also popping up on platforms like RunningHub.
Why it matters
For creators restoring archival films, upscaling 3D animation, or handling video mastering, Spark VSSR replaces blind guesswork with precise control. Getting a clean output across hundreds of frames just by perfecting two or three keyframes is a huge workflow win.
Just as important is the decision to open up the entire 42.2 GB training and inference stack. In our view, while the steep GPU requirement keeps casual users away, having full access to weights and training code gives serious developers an exceptionally solid foundation to build on.
Sources (1)
- github.com(opens in a new tab)Originalgithub.com