RFD 1160: SkinTokens fits at every precision

Problem

SkinTokens auto-rig is 0.5 B parameters: 1.0 GB at bf16, 0.5 GB at eight bits, 0.28 GB at four. Every precision fits inside 8 GB with room to spare, so memory decides nothing here and the operator question decides everything.

Its input is not an image. Auto-rig consumes a mesh and emits skin weights, and a mesh has no fixed shape: vertex count varies per asset, and the catalog caps it at 210,000.

A dataflow part compiles fixed shapes. A graph whose first dimension is the vertex count of whatever arrives is the wrong shape for this device, whatever its parameter count says.

Decision

Abandoned on 2026-08-28. The accelerator work is scoped to rf-detr keypoint and RFD 1157, and this scored 12 of 25 against RFD 1157’s 18.

Rank SkinTokens low despite the comfortable fit, and say why in the rank: it is variable-topology work, not fixed-resolution work.

Before dismissing it, check one thing. If the model tokenizes the mesh to a fixed token budget, the graph after that point is fixed and the variable part is host-side preprocessing. That would make it a device half in the same shape as rf-detr’s.

Related

RFD 1033 separates the geometric algorithms, which scale with the vertex budget rather than a parameter count. RFD 1131 cuts the graph. RFD 1026 gives the memory.