SV3D takes a single object image as input and outputs multiple new views of that object. These new views and SV3D can then be used to generate a 3D mesh.
When released Stable video spreadingIn our previous release, we highlighted the versatility of our video models across a range of applications. Building on this foundation, we are pleased to release Stable Video 3D. This new model advances the field of 3D technology, offering significantly improved quality and multi-view compared to previously released models. Stable Zero 123and also better than other open source alternatives. Zero 123-XL.
This release comes in two variants:
-
SV3D_u: This variant generates trajectory videos based on a single image input without camera alignment.
-
SV3D_p: Extending the capabilities of SVD3_u, this variant supports both single images and orbital views and can create 3D videos along a specified camera path.
Stabilized Video 3D is now available for commercial use. Stability AI MembershipFor non-commercial use, the model weights can be downloaded below. Hugging Face Read the research paper here.
Benefits of Video Spreading
Stable Video Diffusion By adding camera path adjustment to the image-to-video diffusion model, Stable Video 3D is able to generate multi-view videos of an object. In contrast to the image diffusion model used in Stable Zero123, using a video diffusion model offers significant benefits in generalization and view consistency of the generated output. Furthermore, we leverage this powerful feature in Stable Video 3D to propose an improved 3D optimization that generates arbitrary trajectories around an object. By further implementing these techniques with a separate lighting optimization and a novel mask score distillation sampling loss function, Stable Video 3D is able to reliably output high-quality 3D meshes from a single image input.
View the technical report here More details on the stabilized video 3D model and experimental comparisons can be found here.
Novel view generation
Stable Video 3D brings significant advancements to 3D generation, particularly novel view synthesis (NVS). Unlike traditional approaches that often suffer from limited viewpoints and output inconsistencies, Stable Video 3D can provide consistent views from any angle with better generalization. This capability not only allows for better pose controllability, but also ensures consistent object appearance across multiple views, further enhancing a key aspect of realistic and accurate 3D generation.
