Open Models Almanac
Methodstraining objectivedistillation objective

single mention · training objective

Distillation of specialist depth estimates into autoregressively predicted depth tokens

A VLA is trained to autoregressively predict a depth-token string from an RGB observation, using specialist depth estimates as supervision.

The sources on this page only mention it in passing. The line above is the extraction's one-line gloss. See distillation objective for the methods filed beside it.

Filed under training objective › distillation objective