A lightweight SeedVR tool extracts clean training stills from low-quality video

aoleg77 · reddit · 2026-07-28

A tool that turns low-quality video into training stills for LoRA datasets

The author needed high-quality character stills from an old, low-resolution movie for LoRA training, but simple ffmpeg extraction was not enough. They built a lightweight tool around SeedVR’s temporal super-resolution to detect scenes, pick the sharpest frames in each scene, restore them, and save the results as a dataset.

The project supports segments, uses SeedVR 7B (fp8) in testing, and can be run with a simple command like extract run mymovie.mp4. The output includes restored frames and a gallery stored in a folder next to the source video.

Original post →

More from Multimodal

Multimodal channel →