AI Updates & Free AI Tools You (probably) Don't Know About!

Hello everyone. Welcome back to the Matvid Pro AI YouTube channel. I've got some cool stuff to show off today. We're going to start by covering some brief AI news and then towards the end of the video, I'm going to show you guys some free AI tools you might not have even known existed. So let's go ahead and dive right in. First up folks, I want to make you aware of Vidoo 2.0. This is an AI video generator that often gets overlooked, but it's pretty, pretty darn good. It's competitive, it gets updated quite regularly, and a lot of people kind of seem to miss out on this one. Anyways, they seem to be giving VDU a upgrade, supercharging it a little bit, and it's five days and counting until this 2.0 releases. Now here's the thing about unlike most generators which prioritize real life, you know, something captured on a camera in its training data, Vidoo prioritizes other things like anime generation for example. Example, it's classically known to be pretty decent at animation in comparison to its rivals in the AI video generation world. And you can see for VDU 2.0 they have chosen anime specifically to show off some of its capabilities. And I think this is really cool. I think this is a different approach at the AI generation market. They know a lot of people are trying to generate things that are not necessarily real life. Camera footage of a dog running through a forest or something like that. There might be other features, features with VDO 2.0 that we don't know about. Maybe there will be a specific fine tune available on their website just for animation, or a specific fine tune for non animated works as well. We don't really know. Again, just wanted to make you guys aware in five days this drops. I'm excited to take a look at it and I'm hoping that we can do some really cool AI generated anime style videos with it. Here is some quick non anime style footage that was generated with VDO 2.0. Again, this looks pretty darn good. Not as good as Google's VO2 in my opinion. That model is just ridiculously good. But again, not everyone has access to Google's VO2, including myself. Reports are also coming out that this VDO 2.0 video generation model is really fast at generating. Apparently this clip that you're watching right now only took 10 seconds to generate. But keep in mind the people that have early access to tools like this are obviously on their own dedicated server that's generating these Clips that has quite a bit less traffic flowing onto it than a public video generator. But still guys, sub 10 second video generation times is pretty insane. And even if it's sub minute length generation times on the website once VDU2 is announced, that's still absolutely state of the art and very impressive. Anyways, like I said, the animation really is what excites me the most about this model's release. And this model has been a fan favorite, especially on my Discord server, which you should join by the way. Everyone's really pumped for it. Moving on. Upscaling seems to be taking some precedence right now in the AI space. Tencent just dropped their stable video face restoration tool that turns low quality videos of faces into higher quality ones. So this is obviously an upscaler, an AI video upscaler, but it's got a particular focus on getting the faces correct. So it's not necessarily for general use. It's fine tuned for human faces. And I gotta say, the quality from the examples that are being provided here is pretty stellar. There isn't too much going on in this original input. It's super blurry, not really many pixels to go off of of Albert Einstein's face and it recreates it pretty darn well. Obviously there's some shimmering going on on the higher details here, but it's pretty damn good. And even the colorization here is impressive. This looks like a pretty dang legit Albert Einstein faithfully recreated. I'm especially impressed by how it did the extra portion of skin that's right under his chin here. That's like a super subtle detail that it was able to pick out from the original and then convert over. Here we go with another example. I actually think there's a little bit more detail to grab from the original video in this one, but this video clip has quite a bit more movement going on and a lot more dynamic, dynamic facial expressions. I think it does a pretty decent job. The hair looks shockingly good again, but again there's going to be some shimmering in the finer details. We can see the eye color. It does its absolute best to try to recreate faithfully, but. But even in this original, it's difficult to tell what her eye color is. It kind of goes between brown and blue. In the recreation, the pupils change size a little bit. The reflections aren't perfect. That is what needs to remain consistent because people really do look at the eyes the most when someone is talking. I think the blurrier the original input is, the more impressive the output looks like. This one's Pretty darn incredible. There is a lot of compression going on in her face and the model is able to pick up on the fact that that is compression and shouldn't be recreated in the upscale. That is very impressive to me. But also this is a more high contrast image so it's easier for the model to more forgiving on this person, let's say. Not much detail going on down here in the hair. It's very dark. Minchoi's got plenty of examples if you want to check out the original thread. This thing does a pretty great job. Oh, and for bonus points, this one is open source, which is great to see. This is one of those models that I wouldn't be surprised to see on something like Pinocchio, which is a one click installer for local AI. Next up guys, we have Star AI which is developed by ByteDance, I believe that's the company that owns TikTok, if I'm not mistaken. Anyways, this doesn't just focus on faces, it focuses on the entire image. Now this is full upscaling. Now it's not like this is the first AI based video upscaling tool. There are certainly others, but this one is, you guessed it, open source, which is again more bonus points. Love to see it. Obviously we're not going to see as big of a focus on fully recreating the human face perfectly like we saw in the last one, but it's still quite decent. You can see the original sign here looks quite a bit better on the high definition side of the image, but if we do pause frames, we can see that the sign gets a little bit AI jarbled as the AI tries to hallucinate what that sign is supposed to say. After just a few moments though, it stables out pretty well. And there's an example of the face again not being absolutely perfect with this upscaler, but again this thing's free and open source, so can't complain on two people who are probably in the training data. It seems to actually, actually do a really, really decent job at taking this footage which has been totally destroyed. I mean look at their faces here. Someone call the dermatologist. The upscaler though, brings it back to pretty much like YouTube live stream quality, as if we were actually watching the debates. Pretty amazing. Even the logos and the text down here look pretty good. The American pins though, the American flags don't look great and look I think more like American flags than the original. At some point though, it's like we have such little data to go off of that an upscaling job like this just isn't going to be able to convert that back into an American flag. Now we move on to the non humanoids. And I can say it does a lot better of a job when we don't have real mushy people to try to upscale. I mean, us as humans, we know what a person is supposed to look like and if they're screwed up. But a plane for example, I mean, look at how low quality that original footage is and how much better it looks after the upscale. You can actually see where all the pieces of metal have been riveted together. The reflections are still going from the propellers on the plane and the propellers still look quite good. Also getting that shutter speed motion blur correct on those propellers as well, that's very important to preserve when we're upscaling video. Here is another pretty insane one where it's actually able to tell that that is supposed to be a statue and not a real person. I mean, in the original, at first glance it's kind of difficult to tell if that's a person or not, especially if you're just starting to see the video play. But once it gets upscaled, very clear that this is a statue at a theme park or something. And this is a pretty difficult test. It's got these things hanging off it, obviously. Some rocks, some plants. It did screw up though, I will say, because it's pretty easy for us to tell that that's supposed to be his hand. And once it gets upscaled, it gets turned into like this little face almost. It's hallucinating a little bit. We've also got some food here which get vastly improved by the upscaler as a whole. Does a really great job, especially with the little highlights, a little speckles of reflection from the food that we're watching. Anyways, guys, again, the list goes on By El Cine. This one's also got a completely free demo on Hugging Face that you can check out linked down below. But it's good. I love to see AI being used to bring older, worst quality content back to life. First up, I want to make sure that all of you are aware of Google's Labs effects. It's like their AI lab for you to test things out. They've got image effects, which is basically just Google image in three music effects which actually gets updated more regularly than you'd expect. And then finally, Whisk. I'm going to be showing you guys Whisk because it's pretty darn cool. They do have this screen or we can Start from scratch, but I'll show this off real quick. The style is obviously plushie. We'll upload myself as the subject and then it's going to attempt to turn me into a plushie. And this obviously in some capacity is powered by Google's Imagen 3. But honestly, yeah, these are pretty great outputs. Definitely looks like me if I was a plushie. Getting the details down pretty good from the red shirt that I'm actually wearing right now to the Oura ring, which it has interpreted as just giving me a robotic silver finger. But the hair is pretty accurate, the eyes are accurate. I am a fan. Let's change the style. I went ahead and changed the style to anime and now we can rerun it and oh look, now I'm not really anime matte. But it is getting the General gist here. 2D art. We'll just say that regardless, the whole point of Whisk here is to combine different scenes, styles, subjects, ideas into one final image and you'll end up with some pretty cool stuff. You're whisking or mixing everything together. That's the whole point. It's completely free to try out and use, so I'll link it down below. Now I want to show you a feature in the Meta AI Chat which uses obviously Llama, but it uses some other models as well. This is Meta's Chat GPT competitor and they actually have a feature in here that isn't in Chat GPT at all. And it is actual image editing. I uploaded an image and I said give him a top hat and it was able to essentially Photoshop that top hat right into the image and put it on. My little lemon guy. It looks pretty good too. Now give him a cane and it will start to generate. Boom, there it goes. It gave him a cane. It's right next to him. Deleted the whole leaf though. So yeah, you can do it right inside the chat here, or you can click the little edit button and it will bring you into a specific editing ui, make the hat purple and give him red glasses. Anyways, yeah, if you were to ask Chat GPT to do something like this, it will do its best, but it's not going to be able to capture your original image and then edit on top of it Photoshop style. That's something that yes, the ChatGPT models are capable of. Yes, the Google models are capable of, but they don't give the public access to just yet. Meta AI has it for completely free on their chat bot right now. So if this is something you'd be interested in. I highly suggest this. It's a lot of fun. And again, it is entirely free. Alright, let's upload a photo of me this time. And now we'll do some fun editing. Make candy canes come out of his ears and nose. Is it gonna be able to do this? I highly doubt it. Oops. I can't generate that image. Why not? Is it because it's a face? Put him on the moon. Oh, but it can put me on the mo, which is nice. And you can see it was able to actually crop out my body and even keep the chair in the background as well and kind of put me on the moon. It's pretty funny. Make his eyes 2x the size they are now. Oh, it can't do that. Sometimes it will refuse to do certain tasks. Just the safety of these locked down big tech companies. Huh? Put a picture of a lemon on his shirt. I was able to do that one though. Make him a dog. Oh, can't generate that image. It's just some stuff it will refuse outright. Change the text to say subscribe. Oh, subsolve. Very close. Very close. Yeah, I mean it's more or less just like this little Photoshop style editor that is available for completely free and you can control just by talking to a model in natural language. By the way, Meta also has a bunch of their own demos. They've got a translation one, a quote, quick video generator one, and then also another AI generated music one. So not sure if you guys knew those existed or not, but these areas, the AI demos for Meta, Google or other companies are often updated and they don't post anything about it. So we don't really even know that they're updated until you go check the website. Like Wisk. I had not even seen that before. Anyways, these AI demos usually remain free. The good ones are almost always locked up like Google's VO2 which we've been talking about in this video. But yeah, guys, I hope you learned something new in today's video and I hope it was useful in some way. Thank you so much for watching. I'll see you guys in the next one and goodbye.