Quality problem of MV-Dust3R+'s Resconstruction

[ZhaoLl-code]

I've tried some views in the real world on MV-DUSt3R+,and i found the quality of this resconstruction was worse badly than DUSt3R.The speed of rescontruction was actually faster than DUSt3R...
Maybe some fault I have made caused this result .
I hope someone who got the similar result can contact with me,if you have found the correct way to use MV-Dust3R+

[recordmp3]

For the inference speed, it's as expected that MV-dust3r is faster than mv-dust3r+, which is faster than dust3r.
For the performance, yeah we observed that our method is not generalized well on outdoor/object-centric/human data. When we train another model on them, we will update it and ping you.

[recordmp3]

can you provide your input image?

[ZhaoLl-code]

can you provide your input image?

Thanks for your reply! Here is one of the scenes which i tried

[recordmp3]

We do not trained on object-centric data and sorry that probably our method does not generalize well on that.