|||
It’s time for the Will Smith test. All the video models on Replicate, prompted with “Will Smith eating spaghetti”. Here’s the original version from 2023, and a state-of-the-art version from today. More in thread 🪡
AI video is having its Stable Diffusion moment.
— Replicate (@replicate) December 16, 2024
OpenAI's Sora made everyone realize what is possible, and now there are loads of models that are just as good. Some are even open-source, so you can tinker on them, fine-tune them, and build upon them.
Read more in 🧵 pic.twitter.com/wd6ABZH6CF
Let’s start with this very classy, professional output from minimax/video-01 (aka Hailuo). Surprisingly good! Weird fork physics but right guy, right number of fingers. This is with prompt_optimizer off. Took 3 minutes to generate, cost 50 cents.
The one at the top is also from minimax/video-01, this time with prompt_optimizer turned on. Even better, I think, except for the noodle levitation right at the end. Still a solid result. Again, 3 min, $0.50 to run.
The top open-source model is tencent/hunyuan-video. Here we see a handsome and well-dressed gentleman who is clearly NOT Will Smith, eating spaghetti from a bowl with a spoon. Pretty solid composition but the hands and utensils are sloppy. Took 8 minutes to run, roughly $0.75.
Working our way down the leaderboard, let’s go to the open-source genmoai/mochi-1. This is… not so good. I would even venture to say “disturbing”. Took a solid 11 minutes to generate, costing about $1. Therapist: White Will Smith isn’t real, he can’t eat you. White Will Smith:
Another closed-source model, haiper-ai/haiper-video-2. A little better than the mochi result IMO. This Will does not seem to actually like spaghetti though? And again we see the prompt-enhanced version just yapping. 5 minutes, 30 cents each.
Next in popularity is luma/ray, often called Dream Machine. This bro DEFINITELY likes spaghetti! I like the subtle camera motion, but it’s pretty cartoony. Decent as long as you don’t look at the hands. Fast though, at 40 seconds, and only $0.45.
Finally we have the open source LTX-Video from Lightricks (currently at fofr/ltx-video). This is… uh. Well, it’s a video? I think if you liked the original creepypasta version of this meme, you might enjoy this one. At least it’s cheap, I guess? 12 seconds = ~1 cent for this.
If you liked this thread, retweet and follow me for more stuff like this! If you didn’t like this thread, quote tweet it with your opinion and DM me hate mail! Either way, try generating videos on Replicate today: https://replicate.com/collections/text-to-video
Also shout out to the one and only, the original, the GOAT:
This is getting out of hand!
— Will Smith (@WillSmith2real) February 19, 2024
- Will Smith pic.twitter.com/hHxqB07xC1
