Repository navigation
SceneHarbor restores uneven video shot by shot on a local Vulkan GPU.
- Import a video, compare actual Real-ESRGAN treatments on the same frame, and pick the treatment for each shot.
- Export H.264 MP4 with source frame rate, audio tracks, and chapters. Pause and resume long jobs, or render changed shots after review.
- Export a portable recipe and replay it against a matching source and engine. Use
sceneharbor batchfor unattended queues.
Setup requires Linux x86-64, Python 3.11+, FFmpeg 6+ with libx264, and a Vulkan GPU driver.
uv tool install ./sceneharbor-1.0.0-py3-none-any.whl
sceneharbor install-engine
sceneharbor doctor
sceneharbor serveOpen http://127.0.0.1:8752. Try the attached 12-second showcase.mp4 at 360p. The initial engine install downloads an upstream ZIP of about 45 MiB. The inference workflow runs offline.
On four paired-reference clips, the adaptive selector retained the clean clip and scored 0.889234 mean luma SSIM. Fixed Lanczos scored 0.865286, fixed full-strength AI 0.880293, and fixed Gentle 0.889275. The selector missed a compressed clip where enhancement helped. See the full results and reproduction steps.
The showcase and UI preview use the Sintel teaser © Blender Foundation, CC BY 3.0, edited from the W3C copy. The benchmark notes describe the edits. SceneHarbor source is MIT licensed.