Establish reliable early-exit behavior

Determine whether intermediate states of a trained LoopVAE decoder can be calibrated to produce reliable early-exit reconstructions, potentially through intermediate-output supervision or distillation.

Background

The loop-intervention experiments show that omitting trained passes can substantially degrade reconstruction and can produce severe out-of-range RGB values. Consequently, weight sharing and useful recurrent computation do not by themselves imply that intermediate decoder states are suitable outputs.

The paper identifies intermediate-output supervision or distillation as possible ways to test whether a shared tokenizer can support reliable early exit, but leaves that question unresolved.

References

Intermediate-output supervision or distillation could separately test whether a shared tokenizer can support reliable early exit; the current measurements do not answer that question.

— LoopVAE: Recurrent Depth Across Scales for Visual Tokenization  (2609.11516 - Lu, 10 Sep 2026) in Section 5, paragraph “The high-resolution schedule is a concrete optimization target”