[1/N] Current visual geometry prediction models primarily rely on labeled 3D data. Our CVPR26 paper, Flow3r, allows additionally leveraging unlabeled videos (using flow supervision) for scalable visual geometry learning, enabling accurate multi-view 3D reconstruction in-the-wild.

[2/N] The key question we address is how to incorporate flow supervision. We find that a 'factored' flow prediction module encourages geometric learning (unlike a matching-based predictor) while robustly handling dynamic scenes (unlike explicit geometric projection).

[3/N] Our overall approach is very simple — leveraging a mix of labeled 3D data (via direct supervision) and unlabeled videos (via flow supervision) to train/finetune typical visual geometry prediction networks.

[4/N] We find that: a) our factored flow prediction yields clear gains over matching-based or projection-based alternatives, and that (importantly) b) performance continually improves with more unlabeled data.

[5/N] We also improve an existing large-scale model (Pi3) by jointly finetuning it on 34k labeled scenes and 800k unlabeled videos, resulting in a new state-of-the-art system.

[6/N] Our code, trained models, a large gallery of interactive examples and an online demo for you to try your own samples can be found on our project page:

[N/N] Work led by Zhongxiao (Clara) Cong ( in collaboration with Qitao Zhao and Minsik Jeon. PS: Clara is looking for PhD positions — if you want someone who can lead amazing 3D/4D projects, please reach out to her :)

In-the-wild multi-view reconstruction without 3D labels. That's the jump from research demos to production systems. Checking the code now.