Time-displacement
slit-scan rendering

The core operation reduces to one function evaluated once per output frame:

The vanguard edge (delay = 0) tracks the present. The lagging edge (delay = max_delay) looks furthest into the past. As the clip plays, this temporal rake sweeps through the footage — reading across a single output frame from left to right is equivalent to reading backward through time.

  input frames:   ←──── time ────→
                  [t-N] ... [t-2] [t-1] [t]

  output frame at time t:

    col 0 (vanguard)  ← from frame t      (delay=0)
    col 1             ← from frame t-1    (delay=1)
    col 2             ← from frame t-2    (delay=2)
    ...
    col W-1 (lagging) ← from frame t-(W-1) (delay=W-1)
    

Architecture

The pipeline is fully layered — each stage is independently testable and can be swapped without touching the others.

  ┌──────────┐   ┌───────────┐   ┌───────────────┐   ┌──────────────────────────────┐   ┌──────────┐
  │  Decode  │──▶│ Normalize │──▶│ Frame Buffer  │──▶│           Engine             │──▶│  Encode  │
  │  (PyAV)  │   │  resize   │   │  full · ring  │   │  delay_map · profile · mods  │   │  (PyAV)  │
  └──────────┘   └───────────┘   └───────────────┘   └──────────────────────────────┘   └──────────┘
    

With the 30fps source at 1280px wide, the default max-delay of 1279 frames spans ~42.6 seconds of footage in a single output frame. Reducing to 300 frames (10s) concentrates the distortion; the dancer's body remains more coherent while subtle temporal seams appear at phase transitions.

Wide bands reveal the underlying mechanism: each block is a vertical strip taken from a specific frame, placed side by side. At --slice-width 1 these strips are individual pixel columns and the seams disappear; at wider values they become visible as a temporal mosaic.

How wrap looping works

With --fill black, the lagging bands reference negative frames at the start of the clip — a black fill zone sweeps across the output from right to left. With --fill wrap, those same indices are mapped to index % 992, connecting the end of the clip to its beginning. The output video has no seam and can play indefinitely in an installation loop.

Note: the black band visible in the --fill black examples below is also partly a property of the source footage — the clip opens and closes on near-black frames. This makes the fill zone especially clean here and usefully illustrates the temporal boundary sweeping across the frame.

Modulation patches are specified inline with --mod or in a YAML file with --mod-file. Multiple patches can be stacked on the same destination.

# Oscillate the vanguard position with a 0.1 Hz sine, ±0.4 amplitude
slitscan render 2020.mp4 out.mp4 \
  --profile tent \
  --mod "vanguard=sine:rate=0.1hz,depth=0.4"

# Oscillate max_delay (temporal spread breathes in and out)
slitscan render 2020.mp4 out.mp4 \
  --mod "max_delay=sine:rate=0.25hz,depth=400"

# Both simultaneously — compound modulation
slitscan render 2020.mp4 out.mp4 \
  --profile tent \
  --mod "vanguard=sine:rate=0.1hz,depth=0.4" \
  --mod "max_delay=sine:rate=0.05hz,depth=300"

Modulated delay — temporal spread breathes

Modulated vanguard — the rake shifts

Compound modulation — both destinations

Y-axis modulation

This is the technique of photofinish cameras and 19th-century chronophotography: a fixed slit records motion as it passes, the film advances, and the result is a spatial-temporal composite in which stillness appears sharp and movement creates smeared streaks. Different slit positions reveal different aspects of the subject.

Axis X — vertical slit, time runs horizontally

Each image below is 992 pixels wide (one column per frame) and 720 pixels tall. The slit position selects which vertical slice of the original frame is accumulated. Moving the slit across the dancer's body traces different planes of motion.

slit-position 0.2 — left of frame

Background and edge of the parkway. Sparse motion; parked cars produce faint vertical striations.

slitscan collapse 2020.mp4 out.png --slit-position 0.2

slit through dancer's torso

The dancer's cyclical movement leaves a sinusoidal trace in the image. Each repetition of the dance appears as a periodic waveform.

slitscan collapse 2020.mp4 out.png --slit-position 0.4

slit-position 0.5 — frame center

Centered slit through the dancer's midline. The full 33-second performance visible as a temporal panorama, left to right.

slitscan collapse 2020.mp4 out.png --slit-position 0.5

slit-position 0.5 · direction reverse

Same slit position with time accumulated right-to-left. The performance reads in reverse — the end of the clip appears at the left edge.

slitscan collapse 2020.mp4 out.png --slit-position 0.5 --direction reverse

slit-position 0.8 — right of frame

Edge of frame; cars passing on the right occasionally break through as horizontal streaks.

slitscan collapse 2020.mp4 out.png --slit-position 0.8

wide slit

Wider slit averages multiple columns per frame, softening edge detail and integrating spatial information into the temporal record.

slitscan collapse 2020.mp4 out.png --slit-position 0.5 --slice-width 8

Axis Y — horizontal slit, time runs vertically

Switching to --axis y accumulates horizontal rows instead of columns. The output is 1280 pixels wide and 991 pixels tall (one row per frame). The slit position now selects a height in the frame — feet, waist, chest, or head — revealing how different parts of the body move independently over time.

axis y · slit at feet

High-frequency stepping motion; feet leave rapid oscillating traces.

--axis y --slit-position 0.85

axis y · slit at waist

Lower-frequency lateral hip motion. The waveform period corresponds to the choreography's rhythm.

--axis y --slit-position 0.65

axis y · slit at mid-body

Composite torso motion — a blend of hip sway and arm movement at this height.

--axis y --slit-position 0.5

axis y · slit at upper body

Arms and shoulders: wider lateral range, more complex trajectory than lower body.

--axis y --slit-position 0.25

GIF encoding reduces each frame to an 8-bit palette independently. The dithering introduces a characteristic grain that reads as motion artifact and complements the temporal distortion. Files are large at source resolution — use --resize and --fps for web deployment.

# Output format is inferred from extension — no codec flag needed
slitscan render 2020.mp4 out.gif --fill wrap
slitscan render 2020.mp4 out.gif --fill wrap --resize 640x360 --fps 15

The result: a single vertical slit of the source footage — one specific column of the dancer, picked by --slit-source — is gathered from 1280 different moments in time and laid across the full width of the output frame. Rather than seeing a spatial panorama, you see a temporal panorama of a single spatial point.

When combined with a ramp delay, the slit at position x=640 (center) sampled from frames t, t-1, t-2, ... t-1279 is tiled across the full output width, with each column showing a different moment. The characteristic visual is streaks and halos radiating from motion paths through the fixed slit — the figure's arms sweep ghostly contrails; static background elements produce clean vertical bands.

# Slit at frame center, gathering ~42s of history from that one column
slitscan render 2020.mp4 out.mp4 \
  --slit-source 0.5 \
  --profile ramp \
  --fill wrap

# Slit at left edge of dancer (approx.)
slitscan render 2020.mp4 out.mp4 --slit-source 0.35 --profile ramp --fill wrap

# Tent profile + fixed slit: bilateral temporal fold from one spatial point
slitscan render 2020.mp4 out.mp4 --slit-source 0.5 --profile tent --fill wrap