> As a blind person, AI has changed my life.

Something one doesn't see in news headlines. Happy to see this comment.

One cool feature they added for deaf parents a few years ago was a notification when it detects a baby crying.

Can you share some ways AI has changed your life?

Finally good news about the AI doing something good for the people.

To be fair, the sizes aren't comparable, and for the variant that is comparable, the results aren't that much worse.

These are ~2 years behind state of the art from the looks of it. Still cool that they're releasing anything that's open for researchers to play with, but it's nothing groundbreaking.

I wanted to write exactly the same thing, this reminded me of the Will Smith noodles. The juice glass keeps filling up after the liquid stopped pouring in.

I think you need to go back and rewatch Will Smith eating spaghetti. These examples are far from perfect and probably not the best model right now, but they're far better than you're giving credit for.

As far as I know, this might be the most advanced text-to-video model that has been released? I'm not sure whether the license will qualify as open enough in everyone's eyes, though.

To add things to videos you create with your phone. TikTok and Insta will probably add this soon, but I suppose Apple is trying to provide this feature on “some level”. That means you don’t have to send your video through a social media platform first to creatively edit it (the platforms being the few tools that let you do generative video).

They should really buy Snapchat.

From the paper:

> Datasets. We construct a diverse and high-quality collection of video datasets to train STARFlow-V. Specifically, we leverage the high-quality subset of Panda (Chen et al., 2024b) mixed with an in-house stock video dataset, with a total number of 70M text-video pairs.

This doesn't necessarily mean that it's Wan2.2. People often don't train their own VAEs and just reuse an existing one, because a VAE isn't really what's doing the image generation part.

A little bit more background for those who don't know what a VAE is (I'm simplifying here, so bear with me): it's essentially a model which turns raw RGB images into a something called a "latent space". You can think of it as a fancy "color" space, but on steroids.

There are two main reasons for this: one is to make the model which does the actual useful work more computationally efficient. VAEs usually downscale the spatial dimensions of the images they ingest, so your model now instead of having to process a 1024x1024 image needs to work on only a 256x256 image. (However they often do increase the number of channels to compensate, but I digress.)

The other reason is that, unlike raw RGB space, the latent space is actually a higher level representation of the image.

Training a VAE isn't the most interesting part of image models, and while it is tricky, it's done entirely in an unsupervised manner. You give the VAE an RGB image, have it convert it to latent space, then have it convert it back to RGB, you take a diff between the input RGB image and the output RGB image, and that's the signal you use when training them (in reality it's a little more complex, but, again, I'm simplifying here to make the explanation more clear). So it makes sense to reuse them, and concentrate on the actually interesting parts of an image generation model.

> "VAE: WAN2.2-VAE" so it's just a Wan2.2 edit

No, using the WAN 2.2 VAE does not mean it is a WAN 2.2 edit.

> compressed to 7B.

No, if it was an edit of the WAN model that uses the 2.2 VAE, it would be expanded to 7B, not compressed (the 14B models of WAN 2.2 use the WAN 2.1 VAE, the WAN 2.2 VAE is used by the 5B WAN 2.2 model.)

They used the VAE of WAN like many other models do. For image models you see a lot of them using the flux VAE. Which is perfectly fine, they are released as apache2 and save you time to focus on your transformers architecture...