Reference setsstable
The images
Ten PNGs, each drawn by closed-form integer arithmetic over the pixel grid — no seeded RNG, no fonts, no floating point. --check regenerates them and byte-compares, so the committed bytes are provably the ones the generator draws. They are chosen to spread across the failure modes an image-space preprocessing pipeline actually has, not to be pretty.
| name | Canonical Buffer | bytes | what it is for |
|---|---|---|---|
01-gradient-scene.png | 192×144 | 4,322 | the busy asymmetric scene the CPNP suites use: broad low-frequency structure |
02-text-document.png | 200×256 | 742 | document-shaped: high-contrast thin dark bars on a light page, portrait |
03-high-frequency-texture.png | 160×160 | 59,754 | detail at the grid's Nyquist limit — the sharpest test of how a resize interpolates |
04-low-contrast-field.png | 176×132 | 2,481 | all content inside a 20-level band around mid-grey |
05-saturated-primaries.png | 192×128 | 339 | the corners of the sRGB cube, plus one asymmetric marker |
06-near-black.png | 144×108 | 1,554 | structure inside the bottom 5% of the range |
07-near-white.png | 144×108 | 1,560 | the same, 12 levels below saturation |
08-diagonal-composition.png | 208×128 | 3,247 | non-square, off-centre subject, hard diagonal edge — a centre crop is visible |
09-radial-pattern.png | 176×176 | 26,846 | concentric rings about an off-centre origin, spacing narrowing outwards |
10-photographic-composite.png | 256×192 | 73,781 | photograph-like statistics: sky gradient, sun, ridge silhouette, grain, vignette |
No two items are near-duplicates in any of the three spaces: the largest cosine between two distinct items is 0.916 (siglip2), 0.902 (siglip) and 0.935 (clip). A reference set whose items sat on top of each other would let a tolerance be met by a model that had learned nothing.