Sheesh. Genie 3 from Google DeepMind just dropped, showing how it's capable of generating interactive worlds from prompts, so if you are painting a wall and walk left to right, paint react accordingly. That's... scary. Quality is astounding comparing it to other research papers just from year ago.
In case anyone was wondering, WAN 2.2 in ComfyUI, using the default I2V preset, renders 121 frames at 720p in 1.5 hours on an RTX3090. Cut the resolution in half, and it’s down to 18 minutes. Tweak some settings in KSampler and you’re looking at just 10 minutes. Run that on 5090 and get it in probably around 3 minutes. And there’s still plenty of optimization ahead.
Act Two from Runway comes with some limitations. Objects held in the driving video aren’t transferred into the new style. Every time I generated a video, my hands ended up empty.
Additionally—and this was the biggest disappointment—while we can move our hands closer to or farther from the camera, this doesn’t work with the head at all, meaning you have to stay at a fixed distance from the camera for the entire recording.
You also shouldn’t position your hands perpendicular to the lens, hiding the fingers from view. Act Two tends to freak out in those cases, twisting fingers or generating the wrong number of them.
No doubt these issues will be ironed out in future versions of the tool, which already offers impressive capabilities as it is. Time to test with Aleph! #runway
Timlogy
It landed.
9 months ago | [YT] | 0
View 0 replies
Timlogy
Only about six years late, but hey, great move, Apple!
10 months ago | [YT] | 1
View 0 replies
Timlogy
10 months ago | [YT] | 2
View 1 reply
Timlogy
10 months ago | [YT] | 0
View 0 replies
Timlogy
Sheesh. Genie 3 from Google DeepMind just dropped, showing how it's capable of generating interactive worlds from prompts, so if you are painting a wall and walk left to right, paint react accordingly. That's... scary. Quality is astounding comparing it to other research papers just from year ago.
10 months ago | [YT] | 1
View 0 replies
Timlogy
In case anyone was wondering, WAN 2.2 in ComfyUI, using the default I2V preset, renders 121 frames at 720p in 1.5 hours on an RTX3090. Cut the resolution in half, and it’s down to 18 minutes. Tweak some settings in KSampler and you’re looking at just 10 minutes. Run that on 5090 and get it in probably around 3 minutes. And there’s still plenty of optimization ahead.
10 months ago | [YT] | 2
View 0 replies
Timlogy
Uploaded second part of 3D AI generators! This time you can compare their texture capabilities.
https://www.youtube.com/watch?v=z-9PT...
10 months ago | [YT] | 3
View 0 replies
Timlogy
Act Two from Runway comes with some limitations. Objects held in the driving video aren’t transferred into the new style. Every time I generated a video, my hands ended up empty.
Additionally—and this was the biggest disappointment—while we can move our hands closer to or farther from the camera, this doesn’t work with the head at all, meaning you have to stay at a fixed distance from the camera for the entire recording.
You also shouldn’t position your hands perpendicular to the lens, hiding the fingers from view. Act Two tends to freak out in those cases, twisting fingers or generating the wrong number of them.
No doubt these issues will be ironed out in future versions of the tool, which already offers impressive capabilities as it is. Time to test with Aleph!
#runway
10 months ago | [YT] | 0
View 0 replies