AI

LiFT: Loop Flow Transformers

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 06/10/2026, 13:18:24 UTC.

Briefings

LiFT introduces a generative model architecture that scales computation by looping a Diffusion Transformer core, achieving lower FID scores and fewer parameters than dense baselines. The method uses a continuous depth coordinate to guide inference beyond training depth without retraining, reducing inference FLOPs by 52% while maintaining generation quality.

Key points

• LiFT-L/2 achieves an FID 3.34 points lower than the dense DiT-XL/2 baseline on ImageNet at 256x256 resolution.
• LiFT reduces inference FLOPs by 52% compared to the dense baseline while using approximately 60% fewer parameters.

Context

LiFT represents a new family of looped generative models that scale computation by repeatedly applying a shared Diffusion Transformer core. Unlike standard architectures, LiFT trains each step with a single regression target on a straight path from the initial estimate to a flow-matching target, enabling long rollouts without retraining.

Limitations

The study reports generation improvements on ImageNet at 256x256 resolution but does not provide results for other image resolutions or object types. The specific training depth coordinate used for the continuous path is not disclosed, limiting reproducibility for users requiring exact parameter configurations.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 20/10/2026, 21:21:13 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeibj5qbra6heitgr267ydro4mnvfy2m3gw7mhmkfbbg24takncoa4y