Stability AI presented a system of synthesis of 3D objects from individual images Stable Video 3D

Stability AI presented a system of synthesis of 3D objects from individual images Stable Video 3D

The company Stability AI has released a model for creating 3D videos – Stable Video 3D (SV3D). The model allows you to create short 360 ° videos from one photo. SV3D is based on Stability AI’s previous Stable Video Diffusion model and is available with a Professional subscription for $20 per month. Hugging Face is available for SV3D non-commercial use.

Stable Video 3D AI comes in two variants, SV3D_u and SV3D_p. The first generates 3D objects based on a single input 2D image without adjusting the camera (as a result, objects rotate around an axis).

The SV3D_p option allows you to create a 3D video based on an input image and a 3D object along given camera paths.

The developers note that SV3D uses the Neural Radiance Fields (NeRF) method, which allows generating images from new angles. In addition, the model uses the split lighting method, optimized for the 3D shape and texture of objects.

Stability AI compares the new product with its predecessor, the Stable Zero123 model, which the company released in December. SV3D takes a different approach to creating 3D content. As Zampani explains, Stable Zero123 is based on Stable Diffusion and outputs one image at a time, while Stable Video 3D is based on Stable Video Diffusion and allows multiple new images to be output simultaneously.

According to Stability AI, SV3D provides a clear image at any angle. The company expects that Stable Video 3D will be able to significantly improve the quality of the generation of 3D objects.

Related posts