@anton_lozhkov: It’s really frustrating that @RiversHaveWings @StefanABaumann @Birchlabs et al’s HDiT arch (and NATTEN in general) does…
Summary
A tweet thread advocating for wider adoption of the HDiT architecture and NATTEN, highlighting its significant speed improvements in super-resolution, image restoration, and genome labeling.
View Cached Full Text
Cached at: 07/16/26, 08:15 AM
It’s really frustrating that @RiversHaveWings @StefanABaumann @Birchlabs et al’s HDiT arch (and NATTEN in general) doesn’t have a much wider adoption
- It has been the main workhorse of my super-resolution and image restoration experiments for years.
- It improved the throughput 11x on our genome labeling project at HF.
- It will likely work great for a GroundingDINO-style fine segmentation pipeline I’m trying out.
Seriously everyone, stop killing your model capacity and convergence speed with convolutional layers in UNet transformers, learn from HDiT!
Project page: https://crowsonkb.github.io/hourglass-diffusion-transformers/…
Similar Articles
@ariG23498: I have always admired @stevhliu's work. I consider his technical writeups to be among the best there is. In the latest …
A thread highlighting a technical blog series on how Hugging Face's transformers library loads models efficiently, covering meta device, safetensors, CUDA caching, and more.
@ClementDelangue: The scale of the infra on HF is insane. If you're still hosting models, datasets, agent memory,... in S3 or R2, talk to…
Clement Delangue promotes Hugging Face's infrastructure for hosting models, datasets, and agent memory, claiming it's better, faster, cheaper, and safer than S3 or R2.
@ClementDelangue: The future of biology shouldn’t stay behind black-box APIs. Especially when it touches personal health. Whether you’re …
Hugging Face releases Carbon, an open-source DNA base model that is 275x faster than comparable models, enabling local processing of whole genomes on a single GPU.
Lite3R: A Model-Agnostic Framework for Efficient Feed-Forward 3D Reconstruction
Lite3R is a model-agnostic framework that improves the efficiency of transformer-based 3D reconstruction using sparse linear attention and FP8-aware quantization. It reduces latency and memory usage by up to 2.4x while maintaining geometric accuracy on backbones like VGGT and DA3-Large.
@AryaTschand: I’ve spent most of my internship at nvidia architecture research thinking about how to frontier llms should inference o…
An NVIDIA intern shares insights from research on how frontier LLMs should perform inference on heterogeneous systems, with a thread containing TLDR and mini experiments.