Question about Samplers

PixArt-alpha / PixArt-sigma

PixArt-Σ: Weak-to-Strong Training of Diffusion Transformer for 4K Text-to-Image Generation

https://pixart-alpha.github.io/PixArt-sigma-project/

GNU Affero General Public License v3.0

1.44k stars 68 forks source link

Question about Samplers #65

Closed YanzuoLu closed 2 months ago

YanzuoLu commented 2 months ago

Hi, thanks for your great work! I'm curious why the sampling of pixart models must use DPMSolver? I notice a significant quality drop when use DDIM even the steps are increased to 50. To best of my knowledge, the convergence should be the same? Below shows the diff between DPMSolver-20-Step and DDIM-50-Step. 1 girl, by Sorayama-dpmsolver20 1 girl, by Sorayama-ddim50

lawrence-cj commented 2 months ago

It's the advancement of DPMSolver against DDIM. Not just for PixArt I think? You could try DDIMStep=100. Then they may be closer. BTW, for ODE solver, you can also try SASolver which is also our default setting.

YanzuoLu commented 2 months ago

Okay, it's quite strange since increasing the inference steps to 100 does not bring much difference from the 50-Step one. While the advancement of DPMSolver against DDIM in UNet-based SD models is almost invisible. Anyway thanks for your suggestions!