Original image → filter → moving content
The unprocessed image stays fixed. Inside a stationary 384 px circular window, the Gabor-filtered RGB version of the same image translates at 24 px/s. The window and its 24 px cosine edge do not move. No separate object is inserted.
RGB bandpass + DC: each linear RGB channel is convolved with the signed Gabor kernel; only constant mean RGB is restored. A filter selects existing image structure, so it need not produce obvious regular stripes. Color lowpass + bandpass additionally retains a Gaussian lowpass (σ = 4 px), a broadband control with extra motion cues.
SF is 4, 16 or 32 cycles / 384 px. Contrast scales deviations around fixed mean RGB; this is not Michelson contrast. Existing source gains are fixed across SF and both branches, so filtered RMS may differ across SF. The finite window broadens the displayed spectrum. These previews are not an isolated single-frequency or human-calibrated experiment.
Classic reference has the same SF, speed, window and duration on gray. Its mean and RMS contrast are not matched to the natural image. Browser previews use existing quantized filter atlases, browser-resized backgrounds and linear RGB compositing. The rightmost Earlier classic retains the 7 Oct sample and its own independent player; new controls do not change its parameters. No new model results are attached.
Same-source fixed-aperture v3 · 512 × 512 · 60 fps reference clock · 2 s. Current display timing is not a calibrated exposure.
Related methods: Perfetto et al. (2020) distinguishes spatial filtering and contrast normalization in natural images; Wallis et al. (2015) modifies selected frequency-band contrast in natural movies. Our RGB direction task is an adaptation, not a reproduction of those behavioral protocols.
Original atlas protocol & measurements · Earlier 42 encoded aperture samples · Earlier eight moving-object samples & QA · RGB multiplication branch