🚀 Our paper **MDMP** has been accepted at CVPR’25 - HuMoGen 🚀. We propose a multi-modal diffusion model that fuses textual action descriptions and 3D skeletal data to generate long-term human motion predictions, with interpretable uncertainty — paving the way for safer and

Such a cool feature!! Being able to import a 3d reconstructed scene through a Mesh format in Blender is a game changer. I guess the next step is 4D dynamic scenes with deformable meshes.

RT @eustachelb: Transformers support is coming today! 🤗

Inspired by:.• Gen3C → • Diffusion-as-Shader →

5. Rendering: Finally, we feed the depth video into a video diffusion model – basically using it as the renderer. This neural renderer “paints over” each coarse frame with higher quality details, adding realistic textures and lighting. Because it’s guided by the depth and the.

4. Depth Video (Coarse Render): From the animated 3D scene, we add a precise camera trajectory and render a depth map for each frame. This “depth video” looks rough, but it encodes the exact 3D structure and motion. These depth frames act as a blueprint, providing geometric.

3. Animation: To bring the scene to life we rig and animate the 3D models. Essentially, we leverage standard 3D animation tools and open source ML rigging & animation models to choreograph how everything moves. For example, we can make a character walk or have a car drive across.

2. 3D Scene Reconstruction: We use images or prompts to generate the 3D scene. Using single-view 3D reconstruction, we can generate a rough 3D mesh of the entire environment and its main assets. Foreground elements (like characters or cars) can be extracted into separate 3D.

1. The building point of this project is the observation that using accurate depth-videos generated from 3D reconstructed scenes allows to provide rich and structured conditioning signals that makes the process of video generation more "3D-aware" without being over-constraining.

Controllable 3D-aware Video Generation: I'm developing a pipeline that builds a 3D scene and uses a controllable video model as a renderer. Why? To improve the video generation controllability over the scene composition as well as the camera motion. The goal is to use the

RT @Michael_J_Black: It's clear that video diffusion models know a lot about the 3D world, material properties, and lighting. The trick is….

Presented our poster at CVPR 2025 / HuMoGen this week in Nashville!.Amazing to share ideas with the community and see our work in motion🤸. Thanks to the @humogen11384 organizers for making it all happen. 🔗 #CVPR2025 #HuMoGen #AI #MotionPrediction

RT @forgegfx: Open Sourcing Forge: 3D Gaussian splat rendering for web developers!. 3DGS has become a dominant paradigm for differentiable….

RT @DeemosTech: 🚨 Paper Alert. Our recent breakthrough CAST: Component-Aligned 3D Scene Reconstruction from an RGB Image has been accepted….

RT @OdinLovis: I am really super happy to show you my research that transform 3D volumetric capture of man capture with @kartel_ai and with….

very interesting article on Spatial Intelligence and 3D Awareness through Point Tracking: From my experience with current Point Tracking technics, the main issue that I've run into for generating a video by animating an image conditioned on Point

I have been working on testing it in Blender and Nuke and the estimations of 3d camera trajectories of DPVO+SLAM are pretty impressive, could be very useful for matchmoving.

RT @akanazawa: Exciting news! MegaSAM code is out🔥 & the updated Shape of Motion results with MegaSAM are really impressive! A year ago I d….

RT @sidahuj: 🛠️ Blender MCP update:. Bring in high-quality assets to Blender through just prompts, thanks to an integration with @polyhave….