Where videoskillet sits
Several projects make video look like it went through composite, tape and a CRT, and they differ in what they operate on. videoskillet simulates one signal, live, so every fault lands on the same waveform.
The neighbours#
ntsc-rs#
ntsc-rs shares the premise of simulating the path rather than drawing the look. It ships standalone, in a browser, and as AE / Premiere / OpenFX plugins, and its multithreaded SIMD Rust runs in real time well above NTSC resolution. videoskillet is fixed to the NTSC raster and has no plugin yet (the FAQ).
BENDR#
BENDR is the closest neighbour: a live browser tool with four channels, a reorderable chain on each, three mix buses, keys and wipes, all in one HTML file a phone will run. Each fault — chroma bleed, rainbow fringing, dot crawl, ringing, line-by-line sync tears — is a slider drawn onto the picture, independent of the others, so the stages reorder freely.
videoskillet builds the signal: a picture becomes a composite waveform, the model damages that waveform, and a model of a TV decodes it back. Dot crawl and rainbow fringing are then leftovers of a decoder that could not separate colour from brightness cleanly, so every fault on the signal interacts with every other one.
vhs-decode / ld-decode#
vhs-decode runs the other way: RF tapped off a working deck's head amp, captured with a CX card or a Domesday Duplicator and decoded in software — VHS, SVHS, U-Matic, Betamax, Video8 — out to timebase-corrected luma and chroma. It supplies real signals, so a claim made here can be checked against one.
Blargg's filters and the RetroArch CRT shaders#
Blargg's filters and the RetroArch shaders solve the console-on-a-period-TV
problem. nes_ntsc and snes_ntsc model composite artifacts for one console's
output, fast and accurate for that case; the RetroArch shaders (crt-royale,
crt-guest-advanced) model the display — mask, scanlines, phosphor, geometry,
glow.
What videoskillet does#
videoskillet is a live instrument. The signal path stays resident on the GPU as compute shaders, so a control change costs one uniform-buffer write. Every stage of the path gets a control, any of them can be driven by an LFO, live audio or a MIDI knob, and two feedback loops run inside the model: a camera aimed at the monitor it feeds, and a mixer patched into itself at signal level. A take renders offline to constant-framerate H.264, and a link carries the look. The features list the rest.
Limitations#
- No plugin, and no timeline. Clips line up in a rundown, and a rendered take carries no audio track (the editor).
- The raster is fixed at 910×525 samples, 754×480 active, so a 4K source is sampled down to NTSC resolution.
- A take is only reproducible from clips. Offline renders of one take come out identical; a camera, a screen share or the mic is live capture.
- It needs a WebGPU browser. On Linux, Firefox Nightly or Chrome.
- The model is progressive 525/60 rather than interlaced at field rate, the largest remaining authenticity gap (the architecture).
All free and open source. Written from the other projects' own documentation and source, not from benchmarks run here. If something is out of date or unfair, please open an issue.