Everything is a Remix

View Original

OpenAI’s Sora: Toy or Tool?

All new technologies begin as toys. Cars, cameras, and personal computers were all once gadgets of unclear utility. But truly important technologies eventually cross the threshold from toy to tool. 

Whenever you start working with an exciting new technology, you first need to know whether it's a toy or a tool.

If it’s a toy, you’re not expecting a lot. You’re looking to experiment, learn, and have fun.

If it’s a tool, the bar is way higher. You expect to get actual work done. 

OpenAI’s Sora has been public for a while now. I’ve used it a lot and I’ve seen hundreds of clips generated by others. Is Sora a toy or a tool? The verdict is clear.

Video is like really hard

First let me say this: the mountain Sora is climbing is very, very high and very, very steep. Video is immensely complicated. Images are far less complex and yet AI image generators still regularly give us stuff like this.

AI image generators have come a long way and yet… (Generated by Leonardo.Ai)

AI video generation has the all same challenges as image generation… multiplied by at least a hundred.

Video generation requires creating hundreds of sequential images which must:
1. Look good
2. Look right
3. Flow together coherently
4. Convey your intent

That’s a lot of balls of juggle. How does Sora fare?

See this content in the original post

Sora is amazing

Sora is mind-bendingly amazing. If I had seen synthetic clips like the ones below ten years ago, I might have thought aliens created them.

If you look closely, you might find rendering issues, but overall these look good and could possibly be used in projects.

Now check this one out. 

It also looks pretty good, but look at this bizarre still. 

If you watch the clip above closely, it’s riddled with morphing glitches like that. And that’s just the start of the issues with Sora.

Sora is also awful

Clips like the first ones I showed are a rarity. Far more typical are clips like these.

Using Sora often feels like throwing spaghetti at the wall. I tried to create that in Sora and got this. Actually, this one of the better ones.

The majority of clips Sora creates are unusable. They range from looking a little weird to totally surreal.

Now some of you might be thinking, Kirby, do you love slow motion or something? Why are you showing so much slow motion?

None of these prompts mention slow-motion. This is a weird quirk of current AI video generators. They all seem to create slow-motion video by default. That’s a big issue.

And the verdict is…

The verdict is pretty clear, right? Awesome as it is, Sora is firmly in the toy category. The quality of its output is just too erratic for it to qualify as a tool. Not only is Sora a toy, but it’s a very expensive toy: the full version is $200 per month.

But it’s still early in the game. What does the future hold?

My guess is that Sora and other video generators will become much more reliable and convincing. Even more importantly, they’ll get integrated into video-specific applications like After Effects, Davinci Resolve, and Capcut, which will enable us to work around the glitches. Video generators will be capable of generating b-roll, backgrounds, and support elements.

Will video generators become good enough to wholly produce immersive films? I won’t be waiting for that. Again, the mountain Sora is climbing is very tall, very steep.

For now, Sora is a fun toy. But if you’re serious about leading-edge video production, dive in and enjoy.