IAAILoopVAE
LoopVAE: Recurrent Depth Across Scales for Visual Tokenization
Reuses a scale- and loop-conditioned core within and across spatial scales of a hierarchical visual tokenizer, keeping only resolution-changing transitions independent.
Inside the method
[ 4-block core (scale, loop conditioned) ] × 28 applications, transitions untied
Simplified conceptual schematic. Consult the paper for the complete architecture.
- Recurrence family
- Core recurrence
- Depth control
- Fixed recurrence across scales
- KV / state strategy
- Not applicable (tokenizer)
Reading note
The 29M convolutional model reaches 0.28 rFID and 32.54 dB PSNR on ImageNet-256 with about 65% fewer parameters than 84M reference VAEs. Fewer stored weights mean more arithmetic and longer runtime in tested settings.
Sources checked 2026-10-28. This catalog does not imply independent reproduction.
Related works
Cite this work
@misc{loopvae2026,
title = {LoopVAE: Recurrent Depth Across Scales for Visual Tokenization},
author = {Zhiying Lu},
year = {2026},
eprint = {2609.11516},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2609.11516}
}