LiFT: Loop Flow Transformers
LiFT loops a shared Diffusion Transformer and lowers ImageNet FID while cutting parameters and FLOPs.
Loop Flow Transformers (LiFT) repeatedly apply a shared Diffusion Transformer core, training each step to regress a point on a straight path toward the flow-matching target. A continuous depth coordinate lets a trained model loop beyond its training depth without retraining or early exits. On ImageNet at 256x256, LiFT-L/2 reports an FID 3.34 points lower than a dense DiT-XL/2 baseline while using about 60% fewer parameters, 32% fewer training FLOPs, and 52% fewer inference FLOPs.
- LiFT repeatedly applies one shared Diffusion Transformer core.
- Continuous depth allows looping past training depth without retraining.
- LiFT-L/2 FID is 3.34 points below dense DiT-XL/2.
- About 60% fewer parameters and 52% fewer inference FLOPs.
From the original article
1 min read at huggingface.co
We introduce Loop Flow Transformers (LiFT), a family of looped generative models that scales computation by repeatedly applying a shared Diffusion Transformer (DiT) core, with only light changes to the standard architecture.
By Mohammad Mahdi Derakhshani, Pedro M. P. Curvo, Gertjan J. Burghouts · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.