Profile 0 (classic) in RGBV3D resembles a PNG mental model: spatial prediction across the raster, then Deflate on the residual stream. Profile 1 (vector-v2) tries to exploit directional coherence in RGB space with a codebook plus exact residuals. Neither profile promises a win on every file.
Where classic often competes: flat UI screenshots with large solid regions, simple diagrams, and some gradient banners. PNG’s mature predictors are hard to beat on arbitrary photos, but classic can land within a few percent when prediction matches the panel structure.
Where vector-v2 sometimes wins: synthetic images where hue changes smoothly and magnitude is stable—think dark-mode app chrome or vector-like artwork rasterized at exact dimensions. The benchmark table at /rgbv3d/benchmark/ lists lab samples; the UI screenshot row may show vector-v2 below PNG, while the film-grain photo row may show both RGBV3D profiles above PNG size.
Where we lose—honestly: high ISO noise, scanned paper, and dithered textures. Noise defeats direction clustering; residuals inflate. Film grain can make vector-v2 larger than raw RGB because metadata and codebook overhead dominate. If your archive is mostly camera JPEGs, PNG or AVIF lossless may remain smaller; RGBV3D is still useful when you need provable round-trip through our tooling chain.
Choosing a profile: use classic when your assets look like “PNG-shaped” data; try vector-v2 on UI captures and solid-gradient art. Always keep the PNG original until you have measured on your corpus. The web encoder shows ratio vs uncompressed RGB immediately after encode—use that as a hint, not a contract.
We publish losses because reviewers and engineers can smell fixed “60% smaller” claims. The format page describes block modes; the benchmark page shows numbers. That pairing is the site’s differentiated content— not another generic compressor landing page.
2026-06-15