Summary · Chapter 16

Tokenization beyond patches

Five families of “fewer, smarter tokens”.

FamilyExampleKey factVersus this project
Merge or pruneToMeAbout 2x throughput, 0.2-0.3% drop (ViT-L at 512 px)No geometry or read-out
Content patchesAPT, SPiTAPT: +40% (ViT-L), +50% (ViT-H) throughputRectangles or superpixels, one token each
Variable tokenisersFlexTok1 to 256 tokens, nested dropoutAbstract codes, pixel target
Gaussian tokensGViT, GPSTokenGViT-B 76.9% vs ViT-B/16 79.8% (same recipe); GPSToken 128 tokens, rFID 0.65RGB targets, fixed or entropy-split layout
Slots over DINODINOSAURReconstructs frozen DINO featuresNo geometry; objects, not a tiling

Defence answer. “GViT is the closest: 2D Gaussians as tokens predicted feed-forward. Mine targets a frozen DINOv3 feature field, solves features in closed form, serves any count from 16 to 256, and places Gaussians by unrolled clustering and Adam. DINOSAUR shares the recipe of frozen features and soft competition, without geometry.”

Open item: equal-count baselines (pooling, per-image k-means, token merging) are not yet run.