search

willisma / SiT

Official PyTorch Implementation of "SiT: Exploring Flow and Diffusion-based Generative Models with Scalable Interpolant Transformers"
https://scalable-interpolant.github.io/
MIT License
589 stars 27 forks source link

Has anyone encounter that using SDE will produce nan vector while the results by using ODE is correct? #22

Open xmhGit opened 1 month ago

willisma commented 2 weeks ago

This could be a numerical stability issue, and you could try adjusting the start and end interval for the SDE integration depending on the specific model prediction and interpolant you are using.

xmhGit commented 1 week ago

This could be a numerical stability issue, and you could try adjusting the start and end interval for the SDE integration depending on the specific model prediction and interpolant you are using.

Thanks for your reply. I will try that!

xmhGit commented 1 week ago

I would be grateful if you could provide some references or links about setting these specific intervals.