1.00

v1 is a raw win32 API app with an entirely custom-drawn UI. unlike old version which was purely a GUI shell that calls out its external core, v1 embeds the optimization pipeline directly. far more ambitious than its predecessors, v1 introduces new features that make it an image editor specialized in image optimization

viewer

pinga has now a slider to compare the original against the optimized version side by side. a set of keyboard shortcuts keeps you in flow. it displays the image with adjustable zoom (mouse wheel, automatic fit-to-window via fit zoom), a pixel grid shown automatically past a certain zoom level — pixel-art editor style rendering. beyond that, the viewer also lets you edit pixels directly, giving you hands-on control over the image itself

lossy compression

• auto — analyses the source image (grayscale content, transparency, color count, flat vs. gradient vs. picture vs. text regions) to classify it, then automatically delegates to either the palettizer (for palette-friendly, flat-color images) or the filter method (for photographic/gradient images) — the user does not pick a method, pinga picks it based on image content

• filter — runs a predicter pass that reworks the PNG filter/prediction data at a strength driven by the quality percentage, trading fidelity for a smaller residual stream without reducing the number of colors

• palettizer — converts the image to an indexed palette: quality maps to the dithering strength while the number of colors would be editable through a dedicated action (and window)

these features stay hidden by default and only appear once user have explicitly turned them on in the settings

• rdo — rate-distortion optimization passes the quality value (1–100) directly to an rdo encoder that adjusts pixel values to make the subsequent deflate compression more efficient

• near-lossless — runs a quantizer pass driven by the quality percentage that nudges pixel values toward more compressible patterns while staying visually close to the original

• color quantizer — palette-reduction mode where the user directly sets the target color count (257 to 8192) and a dithering strength, rather than deriving them from the quality slider, giving finer control over the palette/size trade-off

• denoiser — applies a noise-reduction pass first then re-runs the filter/predicter pass on the denoised result — useful for photographic images where noise inflates the compressed size

palettizer & quantizer

• perceptual space (OKLab) for both clustering and dithering distance
• multistart k-means with weighted centroids initialized from priority-based seeding (population + local density)
• exact kd-tree nearest-neighbor for both clustering assignment and dithering
• alpha as a first-class 4th distance axis with adaptive weighting based on image transparency complexity
• multithreaded k-means and pixel-mapping passes

pending...