AI-driven smooth video-to-video generation

1. A method comprising:

receiving a first video including first frames;

acquiring a text including at least one instruction for transforming the first video;

encoding the text into text embeddings corresponding to the first frames;

encoding the first frames into image latents;

randomly generating initial noise vectors for the image latents;

adding the initial noise vectors to the image latents to obtain noisy image latents;

providing the text to a pretrained motion model to generate animation parameters corresponding to the first frames;

(a) providing the noisy image latents, the text embeddings, and the animation parameters to a first neural network to generate second noise vectors for the image latents;

(b) removing the second noise vectors from the noisy image latents to obtain denoised image latents;

wherein the operations (a) and (b) are repeated a predetermined number of times until the denoised image latents converge to further image latents associated with a true data distribution for latents corresponding to images;

decoding, using a second neural network, the denoised image latents into second frames; and

generating, based on the second frames, a second video, the second video being the first video transformed according to the at least one instruction.

2. The method of claim 1 , wherein the animation parameters are used to modify intermediate features generated by the first neural network during the generation of the second noise vectors.

3. The method of claim 1 , wherein the animation parameters include information concerning positions of an object across the first frames.

4. The method of claim 1 , wherein the animation parameters include information concerning changing a pose of an object across the first frames.

5. The method of claim 1 , wherein the animation parameters include information concerning an action of a character across the first frames.

6. The method of claim 1 , wherein the animation parameters ensure that a difference in positions of an object in adjacent second frames is less than a predetermined threshold.

7. The method of claim 1 , wherein the animation parameters ensure that an action associated with at least one character does not change across the second frames.

8. The method of claim 7 , wherein the action includes one or more of the following: running, jumping, dancing, and playing a musical instrument.

9. The method of claim 1 , wherein the first neural network includes a U-net architecture.

10. A computing device comprising:

a processor; and

a memory storing instructions that, when executed by the processor, configure the computing device to:

receive a first video including first frames;

acquire a text including at least one instruction for transforming the first video;

encode the text into text embeddings corresponding to the first frames;

encode the first frames into image latents;

randomly generate initial noise vectors for the image latents;

add the initial noise vectors to the image latents to obtain noisy image latents;

provide the text to a pretrained motion model to generate animation parameters corresponding to the first frames;

(a) provide the noisy image latents, the text embeddings, and the animation parameters, to a first neural network to generate second noise vectors for the image latents;

(b) remove the second noise vectors from the noisy image latents to obtain denoised image latents;

wherein the operations (a) and (b) are repeated a predetermined number of times until the denoised image latents converge to further image latents associated with a true data distribution for latents corresponding to images;

decode, using a second neural network, the denoised image latents into second frames; and

generate, based on the second frames, a second video, the second video being the first video transformed according to the at least one instruction.

11. The computing device of claim 10 , wherein the animation parameters are used to modify intermediate features generated by the first neural network during the generation of the second noise vectors.

12. The computing device of claim 10 , wherein the animation parameters include information concerning positions of an object across the first frames.

13. The computing device of claim 10 , wherein the animation parameters include information concerning changing a pose of an object across the first frames.

14. The computing device of claim 10 , wherein the animation parameters include information concerning an action of a character across the first frames.

15. The computing device of claim 10 , wherein the animation parameters ensure that a difference in positions of an object in adjacent second frames is less than a predetermined threshold.

16. The computing device of claim 10 , wherein the animation parameters ensure that an action associated with at least one character does not change across the second frames.

17. The computing device of claim 16 , wherein the action includes one or more of the following: running, jumping, dancing, and playing a musical instrument.

18. A non-transitory computer-readable storage medium, the computer-readable storage medium including instructions that, when executed by a computing device, cause the computing device to:

receive a first video including first frames;

acquire a text including at least one instruction for transforming the first video;

encode the text into text embeddings corresponding to the first frames;

encode the first frames into image latents;

randomly generate initial noise vectors for the image latents;

add the initial noise vectors to the image latents to obtain noisy image latents;

provide the text to a pretrained motion model to generate animation parameters corresponding to the first frames;

(a) provide the noisy image latents, the text embeddings, and the animation parameters, to a first neural network to generate second noise vectors for the image latents;

(b) remove the second noise vectors from the noisy image latents to obtain denoised image latents;

wherein the operations (a) and (b) are repeated a predetermined number of times until the denoised image latents converge to further image latents associated with a true data distribution for latents corresponding to images;

decode, using a second neural network, the denoised image latents into second frames; and

generate, based on the second frames, a second video, the second video being the first video transformed according to the at least one instruction.