LiFT loops a shared DiT core to beat dense baseline by 3.34 FID with 60% fewer params
burny_tech · x · 2026-10-21
LiFT (Loop Flow Transformers) is a family of looped generative models that scales computation by repeatedly applying a shared Diffusion Transformer core with only light architectural changes. Each recurrent step is trained with a single regression target—a point on a straight path from the model's initial estimate to the flow-matching target. Because targets are indexed by a continuous depth coordinate, a trained model can loop far beyond its training depth with no retraining or early exits. On ImageNet 256x256, LiFT-L/2 achieves an FID 3.34 points lower than the dense DiT-XL/2 baseline with 60% fewer parameters, 32% fewer training FLOPs, and 52% fewer inference FLOPs.
Related event: LiFT: Looping Shared DiT Core Cuts Parameters 60%, Improves FID(2 posts)→
More from Research
- PyTorch details KTIR and Dataflow Scheduler behind IBM's Spyre accelerator backend — PyTorch · 2026-10-21
- Anthropic: MLP Neurons Are a Stronger Basis for Interpretability Than Assumed — omarsar0 · 2026-10-21
- Microsoft's Self-HPO: Coding Agent Redesigns Its Own Harness, +7 Points on SWE-bench — mark_k · 2026-10-21
- Tim Dettmers: Academia is about to have a renaissance — the best AI work of the next decade happens in university labs — StasBekman · 2026-10-21
- Stanford and Laude release Atlantic-CC, a 17T-token model-filtered Common Crawl for open LLMs — Diyi_Yang · 2026-10-21
- AI saves scientists 7 hours a week, but validation is the new bottleneck — AnneliesGamble · 2026-10-21