I finished most of the Assignment 2 in week 6, which involves generating videos using Pika Labs using images I generated in Midjourney last week as prompts. To summarise my experience with Pika Labs, I want to first discuss some features that I found useful, and then some commands that it was unable to fulfil.
Firstly, I’ve noticed that Pika Labs has recently updated some of its features, the most noticeable one to me being the addition of the area for users to give negative prompts. In my Blog One, I discussed some studies I read about how hallucinations occur due to AI’s inherent weaknesses in processing negative commands. At that point, I thought that this is a drawback for all AI models; but after using Pika Labs, I realise that there are exceptions. In my opinion, Pika Labs have done an exceptional job in understanding my negative prompts, from general ones such as “no morphing, erratic fluctuation in motion”, to specific ones such as “no blinking, large body movements”. As I generated all my videos using images, I noticed that Pika Labs gives better results when I add in specific motion demands. For example, if I use the default setting, it will just zoom in and I won’t get any obvious changes even if I’ve given prompts – it almost looks like I was manually zooming in a still image. But on the other hand, if I tell it to pan left or rotate in a specific way, it will generate a different and better result.
I have also noticed some subtle limitations. Firstly, the resolution is still low, especially if it is compared with image generation tool. Secondly, hallucinations occur quite often especially on human bodies. I’ve gotten many results that gave me two left hands on a human, or six fingers on one hand. Contrastingly, it has been surprisingly successful in generating landscapes, trees, and even aliens – basically anything that is non-human. Considering the inherent challenges that video generating AI are posed as discussed in Blog Four, as well as the significant improvements that Pika Labs has already made, I value and treasure my experiment with it. I have been feeling very grateful for the past two weeks while completing this assignment, as it gave me a chance to have a glance into these amazing tools, and to truly experience and understand the advanced stage AI is already in that not many people know about.
Word count: 404