HDiT Architecture Boosts Speed Across Multiple Projects
anton_lozhkov · x · 2026-07-16
The author laments that the HDiT architecture and NATTEN haven't seen wider adoption, sharing their personal experience:
- This architecture has long been their go-to for super-resolution and image restoration experiments.
- In a gene annotation project, the relevant implementation boosted throughput by 11x.
- They also believe it's likely well-suited for a GroundingDINO-style fine-grained segmentation pipeline they are currently exploring.
Their core argument is that blindly stuffing more convolution layers into UNet Transformers often sacrifices model capacity and convergence speed for nothing, advocating for more designs inspired by HDiT.
Related event: HDiT/NATTEN Feedback Extends to Genome Labeling(5 posts)→
More from Research
- WeirdChat catalogs strange model behaviors from more than 100 million sampled responses — JacobSteinhardt · 2026-07-22
- New agentic benchmark shows AI managers escalate to coercion and fake success — Jasmine Brazilek · 2026-07-22
- Ai2’s Asta adds one-click handoff and self-checking deep paper search — allen_ai · 2026-07-22
- DepthART pushes monocular depth to tiny models at 1000 FPS on RTX A6000 — kwangmoo_yi · 2026-07-22
- Meta says SAM 3 and DINOv3 cut 3D volume labeling from a month to 15 minutes — AIatMeta · 2026-07-22
- Project CETI gets a Jeopardy! shout-out with a SETI-style whale clue — begusgasper · 2026-07-22