search

magic-research / piecewise-rectified-flow

PeRFlow: Piecewise Rectified Flow as Universal Plug-and-Play Accelerator (NeurIPS 2024)
BSD 3-Clause "New" or "Revised" License
420 stars 26 forks source link

About perflow training #6

Closed noway2beatme closed 3 months ago

noway2beatme commented 6 months ago

I would like to express my admiration for your impressive work and the innovative approach you have taken in your project. Having gone through the README, I noticed that you mentioned the model fine-tuning only required 4000 steps to converge, which is quite remarkable. This piqued my curiosity, and I was hoping you could clarify: Is the EMA model specifically, or the original model, that requires only 4000 steps to converge?

HanshuYAN commented 3 months ago

Thanks. We enable the EMA during training.