Isolate the benefit of scale conditioning

Determine the independent contribution of scale conditioning in the LoopVAE shared processing operator, separately from loop conditioning and other architectural components, through controlled experiments.

Background

LoopVAE reuses one fixed-resolution processing core across spatial scales and recurrent loop steps. Learned scale embeddings identify the current resolution, while loop embeddings identify the recurrent step; scale-specific input-injection projections are also used. The paper reports the architecture and its aggregate performance but does not experimentally separate the effect of scale conditioning from the other conditioning and injection mechanisms.

A controlled ablation isolating scale conditioning would clarify whether resolution-aware conditioning is necessary for effective cross-scale parameter sharing or whether the shared core can perform comparably without it.

References

This construction provides an information shortcut; its independent benefit has not yet been isolated experimentally.

— LoopVAE: Recurrent Depth Across Scales for Visual Tokenization  (2609.11516 - Lu, 10 Sep 2026) in Section 2, subsection “Scale-Conditioned Depth Recurrence”

The sharing comparison is a single-run, non-adversarial Transformer experiment; seed variation, identical data order, and separate effects of scale embeddings, loop embeddings, and input injection remain unresolved.

— LoopVAE: Recurrent Depth Across Scales for Visual Tokenization  (2609.11516 - Lu, 10 Sep 2026) in Section 5, paragraph “Evidence boundaries”