Looped DiT: 260M model beats 6.5x larger rival with 4.9x less compute by looping shared blocks
Fit-Ant2238 · reddit · 2026-10-13
A new arXiv paper (2609.40305) proposes Looped Diffusion Transformer: instead of scaling up a text-to-image DiT, the same transformer blocks run multiple times per denoising step, with loop depth adjustable at sampling time. With per-loop deep supervision and self-modulating attention, a 260M looped model beats a 6.5x larger model on several benchmarks using 4.9x less inference compute. The poster ran the MIT-licensed B/16 weights on an M5 Pro: most gains appear by depth 3, and images take 23-24s at depth 1 vs 43-47s at depth 4.
More from Multimodal
- MiniMax-H3 hailed as favorite open release of the year, runs fully local — AdinaYakup · 2026-10-23
- Michal Malewicz experiments with depth maps in his creative workflow — michalmalewicz · 2026-10-23
- One-Shot Tunable Style Transfer on Flux3 Via Reference Latent + Node — Winter_unmuted · 2026-10-23
- PhysVidBench accepted at EMNLP 2026: do video gen models grasp physical commonsense? — aykuterdemml · 2026-10-23
- EMNLP 2026 paper teaches multimodal LLMs to get cartoons via incongruity-resolution supervision — aykuterdemml · 2026-10-23
- A gorgeous Midjourney style: sref code 820221253 for dreamy lavender watercolor looks — umesh_ai · 2026-10-23