GPS · 1.3.29
A real-time WebGL glitch + AI camera FX studio. True CPU Asendorf pixel sort, GPU datamosh, kaleidoscope chained through nine radial UV warps, 40+ generators, on-device MediaPipe Face FX, GEN+CAM live composite, native Save / Load Project, audio-reactive racks, MIDI / VJ control, projector mode.
"Computational vision engine · pipeline: active · fx: multi-layer · audio: reactive"
Also on Google Play (Android) · Desktop unlock code with Android purchase
Artist Statement
I built this because I wanted it.
This is evolving into a single instrument that can datamosh a live cam, run a real Asendorf row-sort, fold it through a kaleidoscope, datamosh that, and paint pixel cells over the top — while hooking up to a projector (projector mode), VJing (MIDI + audio-react), and spitting out a 9:16 GIF.
Glitch Pixel Studio sits in the middle — every effect is a knob you turn while pointing the lens at something. Stack them, automate them, audio-react them, save the patch.
— LOVEBEING
Live Capture · No Post
What's New · v1.3.29 · May 2026
GPU-side dilation via a single uFaceMaskRadius uniform (2.5–38.5 output px) replaces the old CPU separable max-filter. Removed the temporal smear that was costing ~70 ms of perceptual lag. The mask now snaps to the subject every segmenter tick — no lag, no missed frames, full body coverage.
Slider on the INPUT panel. Dial silhouette coverage from baseline body fit up to a 45 px bloom, all on-GPU at zero JS cost.
Typographic finisher post-pass: CHARS · BLOCKS · BRAILLE · SHADES · EDGES. Drop ASCII / Unicode type over the entire composite as a final layer.
BASE (CAM | UPLOAD) and GEN OVERLAY (OFF / FULL / SUBJ / BG) are now separate tile rows with one job each, plus an explicit SWAP FG/BG button.
The intro overlay now hides the camera + segmenter init so there's no dead pause and no gen-only fullscreen flash before FX engage.
Auto-retries startCamera every 2.5 s if the cam is wanted but not live. Segmenter init now retries x5 with backoff for cold-boot WASM races.
Drives true cell size again (chunkier pixels at higher SCALE), not just per-family noise zoom.
Stamp src:* fx:* cam:* seg:Y/n for live state visibility. Color-blind-safe white 2 px borders on active BASE / GEN tiles.
On-device · No cloud · No account · Works in airplane mode
Inside the Rack
Every knob is gated, every uniform is single-source, every session persists.
Not a UV displacement fake — a true CPU row-sort at 256×144 @ 20 Hz, sampled back into TEXTURE5 as uSortTex and gated by a single REALSORT master knob. Luma-keyed (Y << 24 | localIndex) so the streaks read as light/dark runs, not blue-channel noise. AMOUNT, LOW/HIGH, SEGMENT, NOISE, WOBBLE, TEAR, MODE (LINE / SPIRAL / BLOCK / SLICE / HILBERT), INTERVAL gate, ANGLE — all sculpted within the REALSORT envelope.
Twelve dedicated controls: I-FRAME freeze, MOTION bleed, COMPRES (DCT-style block crush), MAP (motion-vector reuse), BLOCK glitch, LIQUID feedback, CHRASH (MPEG smear), FEEDBACK loop, plus SOFT/HARD modes. Linear 0–2 INTENS so the top half of the slider keeps doing work — no dead zone.
KALEIDO chains into nine stacked radial warps — TILE, DROSTE, SPIRAL, YANTRA, MANDALA, ROSETTE, STARFOLD, HEXFOLD, INVERT-SYM. Each is an independent mask-gated UV transform you can stack like a modular synth patch. Combos that don't exist anywhere else.
Bayer dither, halftone, mosaic, voronoi, plasma, ASCII (ertdfgcvb glyph atlas), hex, weave, truchet, stripes, brick, diamond — plus L-systems, reaction-diffusion, boids, particle life, DLA, and a CPU sphere-traced SDF stack. Every generator pipes through the full FX rack.
Front or rear camera streams straight into the shader at full GPU speed via getUserMedia. Plays nicely with anamorphic adapters (Moondog, Sirui, Beastgrip), clip-on macro lenses, gimbal rigs, and external HDMI capture devices presenting as UVC. NATIVE / 9:16 / 1:1 / 16:9 export aspects so your anamorphic squeeze stays intact.
Live mic FFT — bass / beat / level — modulates uSortAmt, uDatamosh, uRGBDrift, generator hue, and rack visibility. MIDI CC mappings for hands-on VJ control, livestream-friendly, single-pass uniform writes so latency stays under one frame.
Dedicated full-screen output with chrome stripped and edge-fade for clean projection mapping. Drive it from the same browser that's running the rack, or fan it out to a second display.
STD / HIGH / ULTRA quality presets, 24/30/60 fps, perfect-loop or full-length, 5/15/30/60 second windows. GIF path uses a quantized palette; video path uses MediaRecorder with optional microphone audio. Hold-the-canvas gesture also records.
Save the patch. Stack racks, automate them, audio-react them, hot-swap from quick-slots. Full session persistence across reloads. Bring up a saved live-set in seconds.
One fragment shader, ~3000 lines of GLSL, ~80 uniforms. Hoisted scratch buffers (Uint32Array(SW=256)) keep GC quiet during sort runs. Auto-quality fallback maps to your device — runs smooth on a Pixel 4 and lights up an M-series rig like a Christmas tree.
Signal Path
Five hops from sensor to export. The shader does the heavy lifting; the CPU sort runs as a sidecar so it never blocks the render loop.
getUserMedia → HTMLVideoElement → glTexImage2D upload at the device's native resolution. Front, rear, or any UVC capture device.
When a generator layer is armed, a procedural canvas composites with the camera feed before the shader pass — so the rack glitches the composite, not just the cam.
One fragment shader runs all enabled racks: pixel sort → datamosh → kaleido + radials → RGBNDR → color → output. ~80 uniforms updated per frame.
When REALSORT > 0, a 256×144 luma-keyed Asendorf row-sort runs at 20 Hz on the CPU and uploads its result to TEXTURE5. The shader blends it back via uSortMix.
MediaRecorder for video (with optional mic), framewise readPixels into a quantized palette for GIF. Output respects your chosen aspect — 9:16 stays 9:16.
Field Notes
Squeeze (1.33× / 1.55× / 1.8×) is preserved end-to-end. Pick the matching export aspect and the GIF / video comes out at the rig's native ratio. Lens flares survive the sort because the luma key reads them as bright runs.
Any UVC-class device — HDMI capture sticks, mirrorless webcam modes, action-cam dongles — appears as a normal camera source via getUserMedia. Use your big sensor, glitch it in the browser.
Projector mode strips chrome and edge-fades for clean mapping. MIDI CC throughout the rack means a Launch Control or Push lights up the whole studio for a live set.
Route a USB audio interface in and the audio rack pulses every uniform. Eurorack to glitch is one cable away. Hold the canvas to record.
The Full Rack
Every effect is its own panel. Every panel is independent, gated, and persists across sessions. Stack them, save patches as presets, assign them to quick-slots. Tap to capture, hold the canvas to record.
Tech Specs
Single-pass shader, on-device AI, zero cloud. Numbers from the shipping v1.3.29 build.
Outputs · straight from the app
Captured live on Android — no post, no filters, no fake demo reel. Just the studio doing what it does.
Built on shoulders
Kim Asendorf's ASDFPixelSort · satyarth/pixelsort · the Gysin / ertdfgcvb ASCII tradition · After Effects Pixel Sorter (Plugin Everything) · classic VGA channel-benders · hg_sdf for the SDF generators · the Akascape pixelsort interval experiments · every modular-synth designer who decided one knob should be the master.
FAQ
The questions that come up over and over.
A real-time WebGL glitch art and AI camera FX studio for Android. It runs a single fragment shader with on-device MediaPipe Face FX, true CPU Asendorf pixel sort, GPU datamosh, kaleido + 9 radial UV warps, 40+ procedural generators, and the GLYPH typographic finisher — all routed through one composite pipeline.
Yes. Everything runs on-device. No cloud, no account, no tracking. It works in airplane mode. Camera frames never leave the phone.
Android 9.0 and up. Auto-quality fallback maps the pipeline to your hardware — runs smooth on a Pixel 4 and lights up an M-series rig like a Christmas tree. iOS is on the roadmap.
Yes. Front and rear cameras stream straight into the shader at full GPU speed via getUserMedia. Plays nicely with Moondog / Sirui / Beastgrip anamorphic adapters, clip-on macros, gimbal rigs, and any UVC-class HDMI capture stick. NATIVE / 9:16 / 1:1 / 16:9 export aspects keep your squeeze intact.
MediaPipe Tasks Vision SelfieSegmenter runs on the GPU at 15–20 Hz with a categoryMask output. v1.3.29 added a faraday-style GPU dilation pipeline (uFaceMaskRadius uniform, 2.5–38.5 output px) that snaps the silhouette to the subject every tick — no temporal smear, no missed frames. FACE mode keeps the person clean while the background glitches; BG mode wraps the person in glitch and leaves the world raw; SWAP FG/BG flips it.
A single knob on the INPUT panel that scales the silhouette outward in output pixels. Left = baseline body fit. Right = ~45 px bloom for full coverage at fast motion. All on-GPU, zero JS cost — moving the slider doesn't drop framerate.
Real. A 256×144 luma-keyed row-sort runs on the CPU at 20 Hz as a sidecar (so it never blocks the render loop) and uploads its result to TEXTURE5. The shader blends it back via uSortMix gated by REALSORT. Streaks read as light/dark luma runs, not blue-channel noise.
Yes. GIF export uses a quantized palette at 24 / 30 / 60 fps with perfect-loop or full-length 5 / 15 / 30 / 60 second windows. Video export uses MediaRecorder with optional microphone audio. Tap to capture a still; hold the canvas to record. Output respects your chosen aspect — 9:16 stays 9:16.
Yes. Native Save / Load Project writes full session snapshots (every rack, every knob) as .gps JSON files to Documents/GlitchPixelStudio/projects/, with a one-tap Share sheet to Drive, Dropbox, Discord, Telegram, anywhere. Quick-slots let you hot-swap presets during a live set.
Yes. Full MIDI CC mapping across the rack so a Launch Control or Push lights up the whole studio. Single-pass uniform writes keep latency under one frame. Projector mode strips chrome and edge-fades for clean projection mapping.
Yes. Live mic FFT (bass / beat / level) modulates uSortAmt, uDatamosh, uRGBDrift, generator hue, and rack visibility. Route a USB audio interface in and the audio rack pulses every uniform — Eurorack to glitch is one cable away.
A live composite that paints a procedural generator over the camera feed before the shader pass — so the rack glitches the composite, not just the cam. Combined with Face FX it stamps generator pattern across just the person (FACE mode) or just the background (BG mode).
Coming soon on Google Play. The current build is v1.3.29.
Runs natively on Android. Uses your camera. Exports glitch art you actually own.
Coming soon on Google PlayA taste of what the app does. The real GPS runs hundreds of agents per frame on your phone — fully on-device.