LiFT Loops a Shared DiT Core: FID 3.34 Lower With 60% Fewer Parameters on ImageNet
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 (DiT) core with only light architectural changes. Each recurrent step trains against a single regression target — a point on the straight path from the model's initial estimate to the flow-matching target — indexed by a continuous depth coordinate, so a trained model can loop far beyond its training depth without retraining or early exits. On ImageNet 256×256, LiFT-L/2 achieves FID 3.34 lower than a dense DiT-XL/2 baseline with 60% fewer parameters, 32% fewer training FLOPs, and 52% fewer inference FLOPs — inference compute can grow independently of parameter count.
Related event: LiFT: Looping Shared DiT Core Cuts Parameters 60%, Improves FID(2 posts)→
More from Multimodal
- Mistral's Voxtral transcribes an 1891 Eiffel phonograph recording at 6.5% WER — IsForAt · 2026-10-21
- Redditor remakes SD clown dance video at 25 steps with motion boost and fixed prompts — roychodraws · 2026-10-21
- Fibo-2, the #1 open image editing model, releases weights on Hugging Face — linoy_tsaban · 2026-10-21
- Nous Research teases generative media plugin for Desktop, already links Krea/Fal — NousResearch · 2026-10-21
- Ex-Brain Games host made a 25-min AI film in 5 days using 'vibe directing' in iMessage — lmoroney · 2026-10-21
- Motion Launches a General Video Agent That Edits Footage and Mimics Any Reference Style — mhdfaran · 2026-10-21