MapAnything V1.1 Release is live! 🚨
@Nik__V__2.1K followers
43,410 views
21 Comments
Comments from the original post on X (Twitter)

@Nik__V__
The key to the new checkpoints is just a better initialization 💡 Splitting DINOv2 ViT-G across both the image encoder and the multi-view transformer!
14 likes

@Nik__V__
We expand the Modular MapAnything framework to now support inference & training of many feed-forward 3D models by using simple wrappers 🎁
20 likes

@Nik__V__
We profile the V1.1 MapA model against the latest SoTA including DA3 & Pi3. MapA is faster! 💨
15 likes

@Nik__V__
We also now include benchmarking comparisons to latest concurrent open source SoTA as of Jan 20th 2026. 🌎 MapAnything is better or competitive to latest models with significantly better metric scale accuracy! 💪 See our 3DV 2026 paper for more details:
14 likes

@Nik__V__
To make benchmarking super easy and accessible we release our WAI format data ⏬
14 likes

@Nik__V__
New improved COLMAP export & input demos with support for adaptive voxelization! For example, below we go from 13 Million points to 300K by removing redundancy - better downstream uses like GSplat 🤗
16 likes

@Nik__V__
Check out our code for more details and to try things out! This marks a freeze on the immediate improvements to the repo. Stay tuned for the next step delta 😉
27 likes

@Nik__V__
For more info on MapAnything:
13 likes

@cavalry1854
Cool work, thanks to FAIR team for updates of their model, easy to miss in constant flow of new releases
2 likes

@Everlier
I'm very curious how it'll behave on the generated videos where perspective is subtly (or often not so subtly) wrong
1 like

@Nik__V__
We have tested this and outcome depends on the magnitude of error in consistency across the video generation. One of our results in the paper is on a generated video
1 like

@Everlier
That's really interesting, I think this might be a part of larger reward function for video models to form better world models. Thank you so much for your time writing this reply!
1 like

@Nik__V__
Yup! This paper does exactly does this:
2 likes

@Everlier
Haha, sorry for wasting event more of your time on sharing the literature with a layman. Thank you!

@Chintu5251
deep learning for 3d is not my type but i found this Interesting.
1 like

@Gibson61350
How's it with with 360 images?
1 like

@Nik__V__
Our model was not trained with them so the camera calibration would not work. However we find the depth estimation to work. Another way of running on 360s is to use a similar logic to MoGe:
2 likes

@davnords
Big!
1 like

@CSProfKGD
💪
1 like

@knaiseris
Is this realtime?
