Stability AI Recently released Stable Video 3D model that creates multiple views from a single image 3D Video.
▲ Photo credit Stability AI
Stable Video 3D consists of two variants, SV3D_u, which generates track video based on a single image input without camera adjustments, and SV3D_p, which extends the functionality of SVD3_u by accommodating a track view and allowing the creation of 3D video along a specified camera path.
Compared to the previous Stable Zero123 model or the open-source alternative Zero123-XL, Stable Video 3D offers a significant improvement in quality, as well as better multiview functionality and more proficient generalization capabilities toMore faithful representation of the input image in three dimensions.
Stability AI says the new model's level of sophistication relies on its cornerstone Stable Video Diffusion model, while Stable Video 3D adds camera path adjustment to generate arbitrary tracks around objects.
Stable Video 3D leverages its multi-view consistency to optimize 3D NeRF and mesh representations to improve the quality of 3D meshes generated directly from new views.
For this purpose Stability AI has devised a new masked fractional distillation sampling loss technique that improves 3D prediction quality. Also its de-entanglement lighting optimization reduces illumination problems and improves shadow quality.
Stability AI states that Stable Video 3D is commercially available through its Stability AI membership subscription ($20 per month for the average individual); and for non-commercial use, it is available on the Hugging Face Download model weights on the platform.