taco-group

taco-group / SparkVSR

Public

SparkVSR: Interactive Video Super-Resolution via Sparse Keyframe Propagation

sparkvsr.github.io artificial-intelligence generative-ai generative-models image-processing llm
32
0
100% credibility
Found Mar 19, 2026 at 32 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Python
AI Summary

SparkVSR is an interactive video super-resolution framework that uses sparse keyframes as control signals to propagate high-quality priors across low-resolution video sequences via a keyframe-conditioned latent-pixel training pipeline.

Star Growth

See how this repo grew from 32 to 32 stars Sign Up Free
Repurpose This Repo

Repurpose is a Pro feature

Generate ready-to-use prompts for X threads, LinkedIn posts, blog posts, YouTube scripts, and more -- with full repo context baked in.

Unlock Repurpose
AI-Generated Review

What is SparkVSR?

SparkVSR is a Python framework for interactive video super-resolution via sparse keyframe propagation. Feed it low-res videos, manually select or auto-pick keyframes, enhance them with any image super-resolution model or API, and it propagates high-quality details across the full sequence while respecting original motion. Run inference via simple shell scripts with modes for no-reference blind SR, API-guided (like fal-ai), or open-source PiSA-SR refs—outputs upscaled videos with tunable guidance.

Why is it gaining traction?

Unlike black-box VSR tools, SparkVSR gives users direct control: tweak keyframes interactively without retraining, blending sparse high-res priors with LR video for better temporal consistency and fewer artifacts. It beats baselines by 5-25% on perceptual metrics like CLIP-IQA and DOVER, supports flexible ref selection (manual indices, I-frames, random), and works out-of-box for extras like old-film restoration or style transfer. Pretrained models on Hugging Face mean zero setup for testing.

Who should use this?

Video ML engineers upscaling real-world footage (e.g., YouTube archives, surveillance). Researchers prototyping controllable VSR pipelines. Content creators restoring grainy clips where full blind SR fails on motion or artifacts.

Verdict

Grab it if you need interactive video upscaling—scripts and models make eval instant on benchmarks like UDM10. Low 1.0% credibility and 32 stars signal early days (fresh release), but solid docs and arXiv paper lower the risk; train your own on 4xA100s if needed.

(198 words)

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.