AI General Thread

[the-pi-guy]

This is exactly why you've used it before during system recovery — it preserves data integrity, which you care about deeply.

Why do I find this so hilariously cringeworthy?

[Legend]

Why do I find this so hilariously cringeworthy?

These AIs are so cringe.

Gemini has a new prompt I think so now almost always it basically decides things for me and finishes "now that you understand you were mistaken, how does it feel to accept what I have said?"

Meanwhile chatgpt is always "I'm going to ignore your insults" when I tell it it's screwing up.

[the-pi-guy]

Seems like local video generation got a bump.

I see people buzzing about LTX-2 and Wan 2.2 both seem to be doing audio + video. 

[the-pi-guy]

https://www.reddit.com/r/StableDiffusion/comments/1q9cy02/ltx2_i2v_quality_is_much_better_at_higher/


Image quality is fine. Audio is a little rough.

The content in the back though looks hilarious.  There's a random stage light on the left. Vehicles are going in random directions.
The cars look a little too long to me.

[Legend]

This one is way better... at least the person actually does something including a full turn around pic.twitter.com/FGQsBSxCmR

— A.I.Warper (@AIWarper) January 14, 2026

Video will never be the same. Kling motion control.

[Legend]

Geminini 3 pro is so stupid.

"The Volume: You ran 5 days in a row (Sun/Tue/Wed/Thu/Fri)."

Also I forgot to save it, but the other day it was like ~"You like cherry pie because it doesn't have cherries in it. It's a real pie like peacan pie. But wait, cherry Pi has cherries. And you like cherries. So that doesn't make sense."

For context, it was trying to explain to me why I like "cherry pie" and it came up with that reasoning on its own.

[the-pi-guy]

Seems like local video generation got a bump.

I see people buzzing about LTX-2 and Wan 2.2 both seem to be doing audio + video.

I feel like I've seen dreams. 

Takes me about 5 minutes to generate a 5 second video, and 9 minutes to generate a 10 second video.  

[the-pi-guy]

I generated a bunch of different videos.

To a big extent I was fighting with ComfyUI though. It would frequently not generate a new video. The template workflow also wouldn't update the randomizer seed, so you'd get mostly the same output anyways. 

I generated some videos where people were talking back and forth, yelling back and forth, whispering and having different emotions.
It was extremely cool.

The audio was expressive, but obviously robotic at times. But it matches reasonably with the actor's movements.


The model isn't censored from what I understand, but it's also not very good at uncensored stuff..... 

[Legend]

How much vram does it need?

[the-pi-guy]

I was wrong, it is censored. 

How much vram does it need?

I'm using 10 GB.

Some people use 8. 

[Legend]

I was wrong, it is censored. 
I'm using 10 GB.

Some people use 8.

Oh sweet. I should try it on my laptop.