zju3dv

Thanks to a tip from the Discord community, we found InfiniSplat v2 is great at panorama-to-3DGS conversion. The feature is now live on the HF Space: huggingface.co And the full workflow is documented on GitHub: github.com/zju3dv/InfiniSplat/blob/main/InfiniSpl… t.co

@simplifyinAI Feel free to try our InfiniSplat v2 model, which is good at turn a 360-degree image into 3D Gaussian Splat: github.com/zju3dv/InfiniSplat/blob/main/InfiniSpl… HF Demo: huggingface.co

360° photo -> Splat oneshot with Infinisplat v2 They don't advertise for it, but the pipeline behind v2 made it very easy to just slice a 360 photo into multiple pieces and let it stitch together by Infinisplat. Works pretty well and better than any other solution I saw so far. :) Props to Infinisplat!

InfiniSplat v2​: our follow-up to InfiniSplat is out. 🤗 Multi-view images in, clean high-fidelity Gaussians out — unbounded free-viewpoint rendering. 👉 Try it now: huggingface.co 💻 Code: zju3dv/InfiniSplat t.co

InfiniSplat v2 released: Analyzing and Improving Feed-Forward 3DGS from Any Views

Excited to introduce SuperNav — bringing CodeX-like agents into the physical world! 🧭 Can accomplish any task in any scene. 📊 Thousands of navigation tasks tested on Habitat-GS, AI2-THOR & HM3D. 🤖 Deployed on a real robot. 👉 Code: zju3dv/SuperNav t.co

@vai_viswanathan Well said. I’ve already forwarded your work to my work group and asked my interns to look into how to incorporate your ideas into the INTACT（zju3dv/INTACT-JEPA%EF%BC%89.

100% AI-Generated Locomotion Animations Using a Video-to-Animation Workflow Idle, walk, sprint, vault, cartwheel, a roll, and a wall run that looks genuinely cursed. None of it was keyframed or pulled from Mixamo. Every motion started as an AI-generated video of a guy on a treadmill. That's a three.js WebGPU playground where the whole animation set comes from 9 clips made with H3 Max on fal (8 image-to-video, 1 extend-video). The motion is lifted out of those videos and fitted onto a mannequin with custom IK and blending. How the pipeline works: - Generate the reference: one image of a person, a short prompt per move, and H3 Max renders the clip - Extract the motion: GVHMR recovers 3D body motion from the video onto an SMPL-X rig - Build the playground: Opus 5.5 put the scene and controller together in about 3 prompts, and cut the parkour clip into vault up, jump gap and fall - Make it speed aware: different variations kick in when standing, walking or running, and longer jumps are stretched procedurally - Extend with text: the cartwheel, roll and wall run came from extending the video moves with UniMate, an open-weight text-to-animation model that runs locally According to the creator, a new set of animations takes around 20 minutes, video generation included. A demo and a free open-source parkour controller are supposed to follow. The rough edges are on screen: cartwheel off a ledge and the character just ragdolls (badly), mantling locks you in place on approach, and the wall run is funny more than usable. Also check licences before shipping, because GVHMR weights and the SMPL-X body model are for non-commercial use. GVHMR: zju3dv/GVHMR UniMate: huggingface.co