IAAIELT
ELT: Elastic Looped Transformers for Visual Generation
Elastic Looped Transformers for image and video generation use weight-shared blocks and Intra-Loop Self Distillation, distilling intermediate-loop students from the maximum-loop teacher in one training step, to obtain a family of any-time models from a single run.
Inside the method
[ Shared generative block ] × r, r ∈ {1 … R} distilled from r = R
Simplified conceptual schematic. Consult the paper for the complete architecture.
- Recurrence family
- Adaptive depth
- Depth control
- Any-time inference at chosen loop count
- KV / state strategy
- Not applicable (generative visual model)
Reading note
With 4× fewer parameters at iso-inference compute, reaches FID 2.0 on class-conditional ImageNet 256×256 and FVD 72.8 on UCF-101.
Sources checked 2026-10-28. This catalog does not imply independent reproduction.
Related works
Cite this work
@misc{elt2026,
title = {ELT: Elastic Looped Transformers for Visual Generation},
author = {Sahil Goyal and Swayam Agrawal and Gautham Govind Anil and Prateek Jain and Sujoy Paul and Aditya Kusupati},
year = {2026},
eprint = {2604.09168},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2604.09168}
}