I’m constantly blown away by new AI video clips, but I always wonder where the real limits are. Well, I just stumbled upon an incredible video where this AI professional decided to find out by throwing the most complex prompts imaginable at all the top models, including Runway Gen-4, Kling, Google’s VEO 2, and Higgsfield.
He systematically tested what works, what’s almost there, and what fails spectacularly. I think this is one of the best reality checks on the state of AI video I’ve seen!
✨ What AI Video Nails Now
The creator started with what’s actually possible today, and the results are stunning. It turns out AI has gotten really good at a few key things:
- 📌 Object Persistence: The expert showed a reflection of a burning building in an eye. The eye closes, then opens, and the reflection is still there and perfectly consistent. This used to be a huge problem, but it’s basically solved!
- 📌 Character Consistency: No more faces morphing into different people mid-shot. The models can hold a character’s appearance and even handle multiple characters interacting.
- 📌 Camera Moves & Style: Basic camera moves like focus pulls look super clean, and the models can handle tons of styles, from anime to photorealistic.
⚙️ What’s Getting Close (The Near Misses)
This is where it gets interesting. The YouTuber then pushed into stuff that’s close but not quite perfect.
One of the coolest things was a “whip pan” camera move. Most models failed, but a new update in Higgsfield actually pulled it off using one of its new presets. The creator also got a model to almost write “hello” in cursive on a chalkboard, and another to knock down dominoes… well, sort of. It kept trying to tip them the wrong way!
😂 Where AI Completely Fails (For Now)
And for my favorite part… the fails! I was laughing out loud at some of these. The expert found that AI video completely falls apart on:
- 🤸 Anything Upside Down: Gymnastics and cartwheels are a hilarious, mangled mess.
- 🏀 Complex Physics: Trying to get a character to shoot a basketball into a hoop was nearly impossible across all models.
- 🧵 Fine, Precise Actions: A simple shot of threading a needle? Nope. Not even close. The models just can’t handle that level of fine motor detail yet.
This summary doesn’t even do it justice. The original creator shows tons of video examples for every single test, plus an awesome showcase of other creators’ work at the end.
For the full deep-dive and to see the wild results for yourself, make sure to watch the original video from this industry pro!