HDiT/NATTEN Feedback Extends to Genome Labeling

Anton Lozhkov recently posted a series of practical notes on HDiT and the related NATTEN architecture, spanning both vision workloads and biological sequence labeling. The posts are notable because they provide two concrete signals: in one genome labeling project, the implementation reportedly improved throughput by 11x; and in eukaryotic ab initio protein-coding region labeling, 1D HDiT was said to perform well on DNA sequences longer than 100,000 bp. Because the posts do not include full experimental context, they are better read as field notes than as a complete benchmark claim.

Key details

According to Lozhkov, HDiT/NATTEN has long been a workhorse architecture in his super-resolution and image restoration experiments. On the genomics side, he specifically said that 1D HDiT looked good for eukaryotic ab initio protein-coding region labeling when handling DNA sequences above 100,000 base pairs.

Separately, he said that in one genome labeling project, the relevant implementation increased throughput by 11x. The posts frame this as evidence that the architecture may be more broadly useful than its current level of adoption suggests.

What remains unclear

The posts do not provide the dataset, baseline, evaluation metrics, or a fuller description of the setup behind either the 11x throughput figure or the long-sequence labeling result. As a result, these statements are currently most useful as author-attributed implementation experience and directional observations, rather than as a fully documented comparative study.

2026-07-16 ~ 2026-07-17 · 5 related posts