mindblowing - but still in the uncanny valley. and I guess it's cute that many of the characters live in a world where AI has caused an apocolypse, but is that really the message they want to lead with?
Lots of activity around Wan lately. It’s nice to see flexible open models make a strong showing against the massively funded closed competitors like OpenAI and Runway.
They're the main open source privacy-preserving video models that VeniceAI has started offering with a convenient frontend. Ovi is one with image to video and there is also Wan 2.1 image to video, Wan 2.2 text to video. Wan 2.5 is available but it's routed in an anonymized fashion to the official provider. They're also much more affordable compared to the routed Kling, Veo, and Sora options.
Heh, I used to work for Nokia's Ovi - basically, gsuite for nokia phones (my group did map search) - the official explanation was "Ovi is Finnish for Door", the internal joke was "Ovi is Hungarian for Kindergarten". I couldn't find any backstory about the name here, though.
This is really amazing, I've been working with AI generation for months and it's amazing how fast separate tools are coming together into one and are usable on your own local machine.
I've been using Ovi for about a week and it's a blast. Like all AI gen, it's a slot machine and even putting in good inputs might lead to bad outputs, but if you run it enough you'll get something good or usable.
I've definitely made many things that look and sound real with both I2V and T2V, albeit T2V tends to look more like 90s tv quality at times, but that also makes it seem more real. If you use Flux SPRO as the image source you can get some pretty realistic looking videos.
I do have a 5090, so it takes about 4 to 5 minutes to make a 5 second clip.
> I do have a 5090, so it takes about 4 to 5 minutes to make a 5 second clip.
what is your setup? took 2 hours for me on 9950x3d with 5090. Any idea what I could be missing? or maybe some other variable is off - i was using default .yml values.
Interesting development, terrible company to have it. I cannot think of a company that has done more to use AI to take advantage of young and lonely people than CAI has.
For I2V You can get a reasonable “draft” out of it in 1-2 minutes with nvidia >= 4070 & sufficient VRAM at 440x440 resolutions. T2V… not as much, still need to be doing something close to trained resolutions, but I still see some quality, more sporadic, at other known-good Wan resolutions.
Also this model seems to benefit noticeably from having both Cuda >= 12.8 and Torch >= 2.8, and separately SageAttention over Flash 2. But I have yet to see any cache threshold with Easy or Tea that doesn’t get a bit postmodern.
Perhaps it's my pre-AutoTune ear, but to me the audio still has hints of perfect pitch and companding. For the former it sounds much like Machine Head in the Invincible series:
Interesting project in the sense of why not see how it works, I'm just still struggling to see the useful usecases of generative audiovisual content. At the moment it just seems to be a nuisance rather than any benefit.
18 comments
[ 0.28 ms ] story [ 41.9 ms ] threadhttps://news.ycombinator.com/item?id=45603435
https://news.ycombinator.com/item?id=45652726
Lots of activity around Wan lately. It’s nice to see flexible open models make a strong showing against the massively funded closed competitors like OpenAI and Runway.
I've been using Ovi for about a week and it's a blast. Like all AI gen, it's a slot machine and even putting in good inputs might lead to bad outputs, but if you run it enough you'll get something good or usable.
I've definitely made many things that look and sound real with both I2V and T2V, albeit T2V tends to look more like 90s tv quality at times, but that also makes it seem more real. If you use Flux SPRO as the image source you can get some pretty realistic looking videos.
I do have a 5090, so it takes about 4 to 5 minutes to make a 5 second clip.
what is your setup? took 2 hours for me on 9950x3d with 5090. Any idea what I could be missing? or maybe some other variable is off - i was using default .yml values.
Also this model seems to benefit noticeably from having both Cuda >= 12.8 and Torch >= 2.8, and separately SageAttention over Flash 2. But I have yet to see any cache threshold with Easy or Tea that doesn’t get a bit postmodern.
https://www.youtube.com/watch?v=DME86-QucsA
Great work all around though.