mirror of
https://github.com/software-mansion/argent.git
synced 2026-09-14 19:27:14 +08:00
Ignacy Łątka
b8d729bb1f
## Problem When two same-aspect screenshots of different resolution are compared (the explicitly supported flow, e.g. a 0.3x saved baseline vs a full-res live capture), `diffPngFiles` calls `normalizeToCommonSize` to downscale both framebuffers to a common size for the pixel diff and sets `imageSize` to those normalized dimensions. The OCR/text pass, however, ran OCR on the original decoded images: baseline OCR bounds came back in baseline-original coordinates and current OCR bounds in current-original coordinates - two different coordinate spaces. `analyzeTextRegions` then compared those bounds directly (boundsDelta / centerDistance / IoU), so identical text at the same logical position was flagged as a half-image `moved` change. The summary compounded it by normalizing the baseline bounds against the smaller normalized `imageSize`, clamping the reported "from" width to 1. A uniform rescale of one framebuffer is not a content change. ### Behavior repro The same "Settings" text rendered at 480x240 vs 240x120 (2:1, same aspect) produced one `kind:"moved"` change with a delta of roughly negative half-image and a "from" normalized width of 1. ## Fix Rescale each image's OCR text bounds (and per-word bounds) into the normalized common coordinate space before the cross-image geometric comparison and the normalized-coordinate reporting. Baseline bounds are multiplied by `(normalizedW / baselineOrigW, normalizedH / baselineOrigH)` and current bounds by `(normalizedW / currentOrigW, normalizedH / currentOrigH)`, so both sets live in the same space the pixel diff already uses. The font-geometry pass is handed the normalized images to match the rescaled bounds, and the top cutoff is derived from the normalized height. The scale factors are 1/1 when the inputs already share dimensions, so same-size comparisons are unchanged. A small pure helper `rescaleTextRegions(regions, scale)` performs the mapping; `analyzeScreenshotTextChanges` accepts optional `baselineRegionScale` / `currentRegionScale` and applies it to the OCR output, and `diffPngFiles` computes the factors from the original vs normalized dimensions. ## Test `packages/tool-server/test/screenshot-diff-ocr-coordinate-space.test.ts` is a pure unit test with no OCR engine dependency - it never shells out to tesseract, python, or PIL. It supplies synthetic OCR regions for "Settings" at the same logical location in the 480x240 baseline space and the 240x120 current space, documents the spurious half-image `moved` change from mismatched spaces, and asserts that after rescaling into the common space the geometry produces no `moved` change and sane normalized coordinates. A pipeline-level case mocks the OCR module and verifies the same through `analyzeScreenshotTextChanges` with per-image scales.