btsmart / splatt3r

Official repository for Splatt3R: Zero-shot Gaussian Splatting from Uncalibrated Image Pairs
Other
507 stars 19 forks source link

Does it support single view feed-forward GS? #19

Closed luminousking closed 1 month ago

btsmart commented 1 month ago

Our model is trained to predict from exactly two images, but we find that in practice if you duplicate your single-view image and feed both copies of the image into the network, then it works well. This is what we do in the demo when you upload one image, and I believe this is similar to how the DUSt3R/MASt3R authors handle the single-view case.

If you are particularly interested in the single-view case, you might want to have a look at Flash3D, which is trained for single-view inference and is based on monocular depth.