Why does the new Miniflux model show great improvement in human structure and motion stability?

[pldlgb]

Hi, I noticed that the new Miniflux model is claimed to show significant improvement in terms of human structure and motion stability. I was wondering, do you have some insights or explanations as to why this improvement occurs?

Looking forward to your response. Thank you!

[feifeiobama]

1. fewer text branch parameters and more visual branch parameters (which was also adopted later in mochi)
2. remove low quality laion dataset in pretraining
3. training from scratch instead of initializing from some pretrained models