RFD 1158: See-Through is the largest model that fits at eight bits
Problem
See-Through layer decomposition is 4.9 B parameters: 9.8 GB at bf16, 4.9 GB at eight bits, 2.70 GB at four. Eight bits fits with 3 GB to spare, which no larger candidate manages.
That headroom is the interesting property. Every model above it is either four-bit-only or sits with nothing left for activations, and today’s failure on a 25 M graph was activation memory rather than weights.
It is four checked-out components, not one: layerdiff, marigold depth, vae and partseg. The 4.9 B figure is RFD 1026’s estimate for the task, and which components it counts is not recorded.
Its output is layers, and no real photograph carries a ground-truth front-hair and back-hair split. So the blinded holdout validates the pose pipeline and not this one.
Decision
Abandoned on 2026-08-28. The accelerator work is scoped to rf-detr keypoint and RFD 1157, and this scored 15 of 25 against RFD 1157’s 18.
It ranked first among the abandoned candidates, on headroom.
LICENSING NOW, NOT FIT. RFD 1166 dropped See-Through entirely: no checkpoint it loads carries a grant, so the headroom above measures a model that cannot be used.