JiT

Back to Basics: Let Denoising Generative Models Denoise

Tianhong Li, Kaiming He · arXiv 2025

Read paper · arXiv:2511.13720

Predict clean data directly and explicitly convert between prediction and loss spaces.

What you will implement

Implement data-to-velocity conversion and the denoising objective with a paper-derived embedded-manifold experiment.

• 17. JiT: Clean Prediction in High Dimensions

  Implement clean-data prediction with velocity-space supervision and compare x/epsilon/v outputs on embedded data.

  Implement make_noisy, prediction_to_velocity, velocity_loss. toy.py provides the training and visualization scaffold.