Two D's for a double dose of deinterlacing.
UPSKAYLEDD is an open-source desktop app and CLI for inspecting, restoring, previewing, and batch-upscaling legacy video with a reproducible pipeline. It is designed to make serious restoration workflows approachable without pretending to be a magic remaster button.
UPSKAYLEDD is a restoration orchestrator, not a magical one-click hallucination box.
The baseline product flow is:
- Drop in files or folders
- Inspect streams, cadence, geometry, and likely source traits
- Present a plain-English analysis summary and a recommended pipeline
- Preview stage-specific before/after results
- Accept sensible defaults or override selected stages
- Run resumable jobs
- Review outputs, manifests, logs, and preview artifacts
UPSKAYLEDD is built around a few simple rules:
- be conservative by default
- let stronger cleanup/upscale paths exist, but make users opt into them intentionally
- fix structural issues before cleanup and upscale
- optimize for truthful previews, not flashy promises
- keep the first-run path simple and the advanced path one layer deeper
- keep GUI and CLI on the same engine
- prefer explicit warnings and fallbacks over silent degradation
UPSKAYLEDD is actively under construction, but it already has a real MVP-shaped vertical slice:
- a shared
AppServiceboundary used by both desktop and CLI - a selected PySide6 desktop shell
- ingest, analysis summary, workspace, queue bar, and dashboard/result review screens
- source-aware stage selection
- preview caching plus exact/approximate preview surfacing
- restart-safe, resumable queue execution
- manifests, logs, and artifact writing
- automated tests covering service and desktop flow behavior
- a Windows-first portable build plus a simple Windows installer build path
This is not production-ready yet, but it is no longer just a bit.
The project is built around:
- Python for orchestration and engine code
- PySide6 for the desktop shell
- VapourSynth for restoration graph assembly
- vs-mlrt for ML runtime and backend bridging
- FFmpeg / ffprobe for inspection, encode, mux, and packaging
- SQLite + JSON artifacts for resumability, manifests, and portable job data
The project should not drift into an FFmpeg-only design that weakens the restoration pipeline model.
The first target user is a technically comfortable desktop user with legacy TV/DVD-era material who wants cleaner, natural-looking 1080-class or higher outputs without hand-building a VapourSynth + FFmpeg workflow.
The motivating source class is:
- SD live-action NTSC/PAL DVD-era content
But the architecture is intended to leave room for:
- SD animation / anime
- progressive HD cleanup
- analog / noisy capture repair
- archive and film-like content
- Windows-first shipping and validation
- Linux considered from the start, not bolted on later
- NVIDIA accelerated where available, but not NVIDIA-locked
- stage-aware interactive previews before long renders
- resumable batch work that survives interruption
- reproducible manifests and readable logs
- hardware optimization as a backend choice, not the product identity
UPSKAYLEDD should not overreach into:
- plugin marketplace design
- distributed rendering
- cloud processing
- interpolation-first workflows
- aggressive one-click “AI remaster” claims
- fake native 4K promises from busted SD sources
- keep tightening desktop UX and preview controls
- beat on real ugly footage harder
- tune safer defaults for live action vs animation
- harden long-run queue, interruption, and resume behavior
- improve packaging and first-run setup on real target machines
Contributions are welcome, especially around:
- source analysis and detection
- restoration preset tuning
- preview and dashboard UX
- cross-platform packaging
- benchmarking and quality validation
- docs that make the tool less intimidating
Start with CONTRIBUTING.md.
Please do not commit:
- giant media files
- private sample clips
- model weights unless there is a very good reason
- generated outputs, previews, caches, or local runtime state
The repo should stay lean. Large media and weights should be fetched or staged outside normal source control.
This repository is released under the Apache License 2.0. See LICENSE.
Yes, the name is a joke.
Yes, the joke is staying.