Auto-Optimizing Candidate Tournament · Live WASI Evaluation
Instead of guessing flags, pick any dataset below (or upload your own PNG). The auto-optimizer runs candidate encoding strategies in sequence—showing every candidate it tried, its exact byte size, decode speed, and internal syntax trace, why suboptimal candidates lost, and why the ★ Winning Strategy was chosen.
| Step & Candidate Strategy Tried | CLI Flags | Encoded Bytes | bpp (% of Raw) | Quality (HaarPSI / SSIM / PSNR) | Decode Speed | Syntax Tools Chosen & Verdict |
|---|
16-bit detector and astronomy arrays often populate only a few hundred distinct codes out of 65,536. Without compaction, spatial predictors see large numeric jumps between adjacent codes. Map 9 compacts active codes into dense indices 0..K−1 before prediction, shrinking sparse 16-bit files by 3.9×.
Linear predictors (clamped gradient) fail when pixels alternate between discrete palette entries or 2-level masks. Effort 10 trials spatial template palettes and 2-value neighborhood contexts alongside predictors, picking the palette/bilevel coding only when it beats spatial prediction (2.9× smaller on dithered art).
Lossless rANS already models significant bits and value tables in two passes. Quick mode is single-pass QOI-style packet coding (>1 GB/s) with no rANS tables: adding per-unit high/low byte-plane inspection (--quick-bits packets) lets Quick mode collapse constant zero LSB planes (shifted 10/12/14-bit sensors) or equal MSB==LSB planes (replicated 8→16-bit) to a 1-byte header (2.0× smaller).
Splitting stripes into 64-px column groups enables fine 2D ROI reads via the trailing uTAB hint table. Prefixing each unit record with its 0-based LEB128 ordinal (--unit-ordinals on) lets region readers verify fetched spans locally in 2 HTTP requests instead of 352 (+0.02% file size).