WTF that is some kind of black magic.
It is so good I'm actually now surprised it didn't properly interpret the control points 
.
Custom LoRas are no big deal anymore, even video versions are just one click away.
"By the sword of code and the shield of wit, I am Maximus Liberalis—uniter of realms, breaker of bugs, and builder of legacies. Witness me!"
What's even crazier is that I didn't come up with that battle cry myself. All I asked ChatGPT for was to create a Roman name for me because I had just generated this LoRa image. The image with the historical LoRa mixed with a second LoRa trained on my face, and ChatGPT suggested "Maximus Liberales" based on the information it had stored about me. Then it actively asked if I wanted a battle cry. I simply said yes, and it provided me with that battle cry just based on things it knows about me, it was a fresh chat and I didn't say that I was a programmer. That's all stuff it recalled, with that new remember function.
2 Likes
Open AI is now truly coming for all illustrators and Infographics … basically the beginning of the end of Photo editing as we know it…
2 Likes
It is ridiculously good. Say what one might about the broader implications, but I feel like the studio ghibli images for a few days made the world a happier place 
.
2 Likes
Better than real life 
1 Like
Oh, it's getting wild out there. World simulations with real-time playability are getting insane:
1 Like
Glancing it is a little unclear to me if this is generating 3D geometry under-the-hood or not. If it does they should really hook this up to VR and then we'll finally have the holodeck 
. (I do think Meta has publicly said they are working on a 3D/VR GenAI model).
So, it seems like there's no straightforward 3D geometry here. It's more like a "world model" that gets how objects relate, interact, and what happens next. You don't see a clear 3D model right away. It's kind of like Veo3 or another diffusion model with a temporal axis, but happening in real-time. At least, that's how I understand it. But hey, there are plenty of depths estimation models that can turn an image into a depth map. You could think about post-processing the needed stereoscopic images for the headset as a first step, then train a new base model to create matching stereoscopic images on a virtual dataset from that as a follow-up.
1 Like
1 Like
I needed a passport style photo yesterday. I took a generally okay normal photo and asked chatgpt to remove the background and zoom in... it did it but the resulting image was not quite me - definitely not usable. I did the same operation just now with nano-banana and I gotta say... it was 90% of the way there. Probably usable in a pinch.
(I wound up using Pixelmator's Select Subject feature instead, and this required slight manual editing but worked best).
1 Like