Features of Kandinsky Neural Network Video Generation

The Kandinsky neural network demonstrates instability when generating video from text descriptions. The model sometimes switches to Russian, and in some cases, traces of the Chinese language appear.

Testing was conducted using the prompt: «One continuous static shot, no cuts. A person looks directly at the camera and counts from one to ten, holding up the corresponding number of fingers as they say each number». Language artifacts are observed during text-to-video processing. Image-to-video generation also reveals specific features of the model.

A distinctive feature of Kandinsky is its selection of character gender traits. Unlike other generators that always create female images for the 'person' prompt, this model tends to generate male faces, including those of Asian appearance.