VectorSpaceLab / OmniGen

OmniGen: Unified Image Generation. https://arxiv.org/pdf/2409.11340
MIT License
2.57k stars 190 forks source link

great job, but why not vae of sd3 #63

Open Robootx opened 1 week ago

Robootx commented 1 week ago

why not use vae of sd3(16 channels) instead of vae sdxl(4 channels), the later has better reconstruction performance,have you tried vae of sd3? thanks

staoxiao commented 1 week ago

@Robootx , when we started this project, SD3 hadn't open-sourced their models.