🔗 Check it out:
👉 Project: m2svid.github.io
📄 Paper: arxiv.org/abs/2505.16565
💻 Code: (coming soon!)

@3dvconf.bsky.social

📊 Results:

✅Higher Quality: Our approach outperforms previous state-of-the-art methods, being ranked best 2.6x more often than the second-place method in user studies.
✅Faster: Runs 6x faster than state-of-the-art competitors.

⚡ Moreover, unlike other methods, we generate a new view without iterative diffusion steps, by training end-to-end and minimizing image space losses.

💡 Our Solution: We solve this by extending Stable Video Diffusion to utilize the input video, warped view (using an off-the-shelf depth model), and disocclusion masks to generate a view from the perspective of the other eye, fixing depth errors and inpainting gaps seamlessly.

🛑 The Problem: Warping standard video to a view from perspective of the other eye is tricky. It creates empty "holes" (disocclusions) and messy depth artifacts where the depth model is inaccurate

Do you want to watch monocular videos in a headset with an immersive 3D experience? We propose M2SVid, a novel architecture that converts standard videos into high-quality, temporally consistent stereo.