nanoFlowMatching - ngromb/nanoFlowMatching-13M

This is a nano implementation of the DiT-LLaMA flow-matching model, trained on CIFAR-10 as part of the CS-503 course.

Model Description

The model consists of a transformer-based architecture that predicts the velocity field for flow matching. It takes as input a noisy image, a timestep, and a class label, and outputs the predicted noise/velocity field. The architecture includes adaptive layer norm modulation (AdaLN) for conditioning on both timestep and class labels.

Usage

You can load this model by adapting the DiT_Llama class from the nanofm.modeling.dit module, to allow pretrained weights to be loaded from this repository.