tapewarp

Make any video look like it was shot on a 1990s camcorder and played back off a worn VHS tape. One Python file, no ML, no cloud, no watermark. Free forever.

This started as an argument with my son. He uses a paid iPhone app for the VHS look on his videos and insisted it couldn't be replicated — "it's not the static overlay, it's how it moves, and sometimes it warps it." He's right about what matters: cheap VHS filters composite a grain texture over the frame and call it a day. The look is in the motion.

So this doesn't overlay anything. Every frame is degraded the way tape actually degrades:

• Wave warp + line jitter — every scanline gets its own horizontal displacement, recomputed every frame: slow sine waves rolling through the image, per-line noise, extra wobble near the bottom of the frame
• Tracking errors — glitch bands that crawl up the frame for a few frames at a time: displaced, noisy, desaturated, washed out (the "smack the VCR" moment)
• Head-switching noise at the bottom edge and a mangled tear band at the top, re-randomized every frame
• Chroma abuse in YIQ space (NTSC's actual color encoding) — heavy horizontal chroma bleed, chroma delayed relative to luma, drifting green/magenta color bands
• Tape ghosting (previous-frame smear), faint dropout smears (real VCRs had dropout compensation, so harsh white slashes are opt-in via --dropouts classic), streaky luma grain, interlace field flicker, lifted blacks, magenta cast, vignette
• Camcorder OSD — blocky PLAY ▶ / SP overlay, running tape counter, and a burned-in date/time stamp in the bottom-right corner, where 90s camcorders actually put it
• Tape-wow audio — pitch wobble, lowpass, hiss

Footage with text on screen is the one case where full strength fights you — title cards and credits get eaten by tracking bands and chroma smear. --wear clean --wobble 0.3 --bleed 0.5 keeps the tape look and leaves the words readable.

Install

You need ffmpeg on your PATH, Python 3.9+, and numpy:

pip install numpy

That's the whole dependency list.

Use

python tapewarp.py your_video.mov

Writes your_video_vhs.mp4. More control:

# how beat-up the tape is
python tapewarp.py clip.mov --wear clean      # fresh tape, one careful owner
python tapewarp.py clip.mov --wear worn       # rented twice a week since 1992 (default)
python tapewarp.py clip.mov --wear trashed    # found in a flooded basement

# period-accurate 4:3 framing (crops like a real camera of the era)
python tapewarp.py clip.mov --ratio 4:3
python tapewarp.py clip.mov --ratio 4:3 --crop-bias 0.3   # slide crop window left of center
python tapewarp.py clip.mov --ratio 4:3-box               # letterbox instead, nothing cropped

# date / time stamp (bottom-right by default, like a real camcorder)
python tapewarp.py clip.mov --date auto --time auto   # parsed from filename or file date
python tapewarp.py clip.mov --date "SEP.7 1998" --time "PM 3:12"
python tapewarp.py clip.mov --date-style numeric      # " 9 7 1998" JVC/Panasonic style
python tapewarp.py clip.mov --stamp-pos bl            # bottom-left (some JVC/Panasonic)
python tapewarp.py clip.mov --date none               # no stamp

# dial the warp back when the frame has text to read
python tapewarp.py clip.mov --wobble 0.3     # gentle drift instead of full wave warp
python tapewarp.py clip.mov --wobble 0       # dead steady (grain and color stay)
python tapewarp.py clip.mov --bleed 0.5      # less chroma smear; colored text stays legible

# extras
python tapewarp.py clip.mov --counter        # running VCR tape counter
python tapewarp.py clip.mov --dropouts classic   # harsh white dropout slashes
python tapewarp.py clip.mov --dropouts off       # no dropout streaks at all
python tapewarp.py clip.mov --no-osd         # image degradation only, no overlay text
python tapewarp.py clip.mov --no-audio-fx    # leave audio untouched
python tapewarp.py clip.mov --seed 7         # different random glitch timing

Filenames like 2020-06-26_13-03-13_000.mov (the default camera-roll pattern) give --date auto --time auto everything they need: JUN.26 2020 / PM 1:03.

How it works

The video is decoded to ~480 scanlines (VHS luma resolution), and each frame passes through a numpy pipeline in YIQ color space. The core trick is a per-scanline displacement field: an array of 480 horizontal offsets, rebuilt every frame from layered sine waves, random jitter, and scheduled glitch events, then applied to luma and chroma separately (chroma gets extra jitter — it really did on tape). Everything else — bleed, ghosting, noise, grade — hangs off that. The result is upscaled back to the source resolution with a soft bilinear pass, which is itself part of the look.

No frame is ever treated the same way twice. That's the whole point.

Credits

• Font: VCR OSD Mono by Riciery Leal (freeware), bundled as vcr_osd_mono.ttf
• Demo footage: Big Buck Bunny © Blender Foundation, CC-BY 3.0
• Inspired by an argument with my son, who was right about the paid filters

License

MIT. Do whatever you want with it.

---

Built by Lanier, an applied AI studio. More free tools at lanierdev.com/tools.