Is your PIV vector field worth reporting?

Drop a PIV image pair. Your browser runs the openpiv pipeline on it - cross-correlation, sub-pixel peaks, validation, outlier replacement, scaling, vorticity - free, nothing uploaded. A paid run then reviews what the field supports or writes the openpiv script that reproduces it.

Examples are synthetic pairs with a known true field, and each has a saved model run - the whole page, free.

Analysis

Frame A (first exposure)

Drop frame A here, or

Frame B (second exposure)

Drop frame B here, or

Drop both files on either box and they are ordered by name. PNG (8/16-bit), uncompressed TIFF (8/16-bit), BMP and JPEG are read as openpiv's tools.imread reads them: greyscale stays as stored, colour goes through openpiv's rgb2gray. Nothing leaves your browser.

Or generate a synthetic pair with a known field

Field Size, px Seed Particles per px Particle diameter, px Noise, grey levels Peak displacement, px Vortex core radius, px Shift angle, degrees

Uniform uses the peak displacement at the angle; shear runs from minus to plus the peak, top to bottom.

PIV settings (openpiv names; defaults are the skill's runner.py)

window_size, px overlap, px search_area_size, px dt, s scaling, px per unit unit sig2noise threshold sig2noise_method subpixel_method correlation_method replace method max_iter kernel_size

drop_invalid (NaN instead of interpolated) global_val, px per frame u min u max v min v max local_median_val threshold, px per frame size

Scaling from a ruler or target in the image: px = mm

Validation thresholds are entered in px per frame and applied in px/s (divided by dt), the units openpiv's u and v are in at that point - the trap the skill warns about.

Drop two frames or generate a pair.

Vector field (free, in your browser, no AI)

Settings tried on this pair

Every PIV run on this pair, newest first, compared with the baseline row. Free; kept until the pair changes or the page reloads.

Settings	Vectors	Flagged	Median s2n	Max \|d\|, px	Truth RMS, px	Browser's read

Background Arrows: valid replaced still NaN

Signal-to-noise

Peak locking

Metrics

Id	Metric	Value	Basis

Flags

By region

Region	Vectors	Flagged	Median s2n	Median \|d\|, px

Flagged vectors

Row, col	x, y (px)	u px	v px	s2n	Why

The skill's runner, same settings

Free: no account, no run.

Notes (optional: the experiment, the flow, what you want to conclude)

Anything to ask (optional)

Load settings .json

Run PIV first to price the review.

Raw model reply

The reply did not parse as the structured result, so it is shown as it arrived.

Your recent runs

What this does, and what it does not

Particle image velocimetry splits two exposures into interrogation windows, cross-correlates each pair of windows and reads the displacement from the correlation peak. The page follows openpiv 0.25.4 and the openpiv agent skill's runner.py step for step: extended_search_area_piv (with an extended search area both windows are intensity-normalised and frame A is masked to its central window; linear correlation zero-pads the FFT), a Gaussian sub-pixel fit, the peak2peak or peak2mean ratio, sig2noise_val and optionally global_val and local_median_val, replace_outliers, scaling.uniform, transform_coordinates and the skill's analyze.py. It was checked against openpiv on 526 image pairs (490 random synthetic pairs over window, overlap, search, correlation, sub-pixel, ratio, validation and replacement settings, and OpenPIV's own six test pairs): 175,631 of 175,637 raw vectors agreed within 0.001 px and the signal-to-noise ratios within 0.01%. The six that did not, and one flag, sit in the normalised path, where openpiv computes in single precision and a near-zero point beside the peak makes the fit sensitive; the page computes in double precision.

openpiv's own quirks are reproduced rather than fixed, and the page says so where they matter: its rgb2gray weights blue by 0.144 instead of 0.114; get_coordinates centres the labels on the image while the windows start at the corner, so each label can sit a few pixels off its window; the linear correlation plane loses one FFT column on the way back. Peak locking, the particle count and the quarter-window check are the page's own diagnostics. One image pair measures displacements in the light sheet - not out-of-plane motion, and not turbulence statistics, which need an ensemble. The paid run reads only what the browser computed and your notes, is told never to compute a new number, and the page checks every number it writes. Derived from the agent skill @k-dense-ai/openpiv (k-dense-ai/scientific-agent-skills; see the notice). No OpenPIV code or data is included.