Parameters for reproducing the FID of the released XL/2

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

662 stars 35 forks source link

Parameters for reproducing the FID of the released XL/2 #18

Closed HanshuYAN closed 2 months ago

HanshuYAN commented 4 months ago

Hi Authors,

Thanks for the great work. I downloaded the released XL/2 ckpt and tried to run the evaluation. The obtained FID value is too high, around 7. Can you share the hyperparameters for evaluation? such as the sampling method, number of steps, cfg scale...

regards, Yan

willisma commented 2 months ago

We used second-order Heun ODE solver with 125 steps (250 NFEs) and 1.5 cfg scale. All training and network hyperparameters follows DiT-XL.