Multi Aspect Ratio Video Rendering 16 9 9 16 1 1

via “custom aspect ratio support with flexible output dimensions”

Generate images from texts. In Russian

Unique: Implements aspect ratio support through VAE decoder dimension adjustment rather than post-processing cropping, preserving semantic coherence across different aspect ratios. Supports both predefined ratios and custom dimensions, providing flexibility without retraining models.

vs others: More efficient than generating square images and cropping because no computational waste on out-of-frame content; more flexible than fixed-aspect-ratio models because single model supports multiple output dimensions.