LiFT expands iterative inference with a shared DiT core

What happened and what to do

The October 6, 2026 publication presents LiFT, which repeatedly applies a shared Diffusion Transformer (DiT) core along a continuous depth path. On ImageNet 256×256, the approach outperforms a dense DiT baseline with about 60% fewer parameters, according to the authors. It offers a way to scale computation at inference time without retraining the model.

For a company, the opportunity is to assess whether iterative inference could reduce the size of a deployed model without compromising results relevant to its product. This could involve testing against internal metrics, monitoring quality and latency, and building an inference API integrated with the existing environment.

How the consultancy can help

Wendelmaques can diagnose the opportunity and requirements, define an evaluation using suitable data and metrics, and implement and operate an inference and monitoring solution on the client’s infrastructure.

Next step

Send a short description of your case to receive a proposal scoped for diagnosis and implementation.