img-collapse

VLMs waste tokens on dead space. Collapse maximal runs of identical adjacent rows/columns into one; keep an index sidecar so any collapsed-coordinate box maps back to the raw original pixels — lossless, no re-rendering.

Stack: python3 + Pillow + numpy (stdlib otherwise). Script is executable, no install.

Usage

skills/img-collapse/img-collapse collapse <in> <out>
    # writes <out> (collapsed PNG) + <out>.map.json, prints orig WxH -> collapsed WxH and bytes before/after

skills/img-collapse/img-collapse crop <orig> <sidecar> --box x0,y0,x1,y1 [--out PATH]
    # box in COLLAPSED coords (inclusive) -> writes the RAW original pixels for that region,
    # prints the original-coordinate box used. Default out: <orig>.crop.png

skills/img-collapse/img-collapse --selftest