1. StabilityAI released a new artificial intelligence 3D video model called Stable Video 3D (SV3D) that creates multi-view 3D meshes from a single image prompt.
2. The model has two variants, one that creates orbital videos based on a single image input and another that allows for single images and orbital views.
3. StabilityAI used a curated subset of the Objaverse dataset for training data, which is available under the CC-BY license for commercial and non-commercial use with proper credit.
StabilityAI has introduced a new artificial intelligence 3D video model called Stable Video 3D (SV3D) that can transform a simple image prompt into a fully animated view of objects. The model, built on Stable Video Diffusion technology, adds depth to video generation by creating multi-view 3D meshes from a single image, maintaining consistency for objects within the video frame. The founder and CEO of StabilityAI, Emad Mostaque, has expressed excitement about the potential for the technology to revolutionize video generation by generating every pixel.
Stable Video 3D builds on previous models like Stable Video Diffusion and Zero123 3D image model, aiming to make the Star Trek Holodeck a reality. The new model offers two variants: SV3D_u, which creates orbital videos based on single image input, and SV3D_p, which allows for the creation of 3D videos along a specified camera path. The technology analyzes the input image to generate multiple views of the object as if a camera was moving around it.
While the full capabilities of Stable Video 3D have not been fully tested, initial sample clips demonstrate the ability to capture objects from various angles and predict non-visible views. The model is currently focused on single objects on a white background, which could benefit companies looking to showcase products with a 360-degree view. The technology has the potential to evolve to handle more complex images and control camera movements in a broader range of scenarios.
StabilityAI has been transparent about the training data used for Stable Video 3D, explaining that it is trained on a curated subset of the Objaverse dataset, a library of annotated 3D objects. The dataset is available under the CC-BY license, allowing users to share, adapt, or remix the material with proper credit. This openness around data provenance sets StabilityAI apart from other AI labs and demonstrates a commitment to ethical and responsible AI development.