Revolutionary! Open Source & Local Video Model STOMPS on VEO 2

I think AI video generation might have just had its stable diffusion moment. Today we're talking about an open source AI video generator that is not only open and free, but also state of the art and can run on consumer grade hardware. We just rolled triples folks. This is WAN 2.1 by Alibaba. So another Chinese video generator here. They seem to be dominating not only the open source AI video generation space, but also the closed source one. This one's also achieved a number one score on V bench here. So it's at the top of the leaderboard, outperforming state of the art open source and commercial models. So this thing is a beast all the way around. No matter how you shape it up, it's a big deal. Even a simple close look at their demo shows off this thing's capabilities. It's when 2.1 we start off with complex motion. We can see woman riding a horse in some long clothing, a dancer on ice, someone breakdancing. And if we play the video we can actually see that it's able to achieve all of these complex, traditionally difficult for AI video generation results without morphing bodies around and mushing body parts together. Incredible. It actually looks like someone breakdancing and not just a compiled mushy mess of horrors beyond our comprehension. And it gets better. We've got a panda on a skateboard. This cool bald eagle fencing competition that's very difficult for video generators. Men's diving also got some options here showing off physics. We're actually cutting apart food and it looks real. There aren't new slices of the food spewing out of it. It just looks normal and natural. And also whilst we're seeing this, you know, a lot of these videos are cinematic quality, high fidelity stuff. More abstract stuff like vines climbing on a window are generated pretty flawlessly. I mean it seems to be a pretty well rounded model altogether. Accomplishing lots of tasks while meeting everything the consumer wants. This is open and free guys. This can be run on a lot of GPUs. Plenty of you at home could probably run this locally for complet completely free. And cherry on top here, text generation and various visual effects to go along with it. So you can do cool, you know, writing intros. Spring growing with vines. I unfortunately cannot read. I assume this is Chinese, but it's doing other languages as well. So it seems to be multilingual in that super cool, very promising. Minimum VRAM requirements here are just under 9 gigabytes. So again that is a lot of consumer GPUs. Here's a quick list at a glance of Nvidia consumer grade GPUs that have over 9 gigabytes of VRAM. So we're going all the way back to the 30 series. Not half bad. And since it's open source, that means anyone can freely and openly modify it. So people can probably make it more efficient as well over time, making the GPU requirements even lower. Here's the hugging face page for this model. If we scroll down we can see there's actually four models to choose from. We've got some larger 14B sized models up to 720p resolution, but AI upscaling has come a long way, so can always upscale these videos to an even higher resolution. Now the model that's going to be using the low amount of VRAM is this 1.3B model. And while it can generate 720p, it's most optimal at 480p. If you've got a real big honking GPU though, you can try to get one of these 14B models to work, which is what I tried to do. I had trouble and we'll talk about why. But before we get into that, I do want to talk about the free options you have to generate with this model. First one is going to be this hugging face space. This is the official web 2.1. Not gonna lie guys, this one is very popular, very congested. You're gonna be waiting in the queue a while for your generations to go through. If you're watching this at a much later date though, and you still want to try it out, this could still be up and generate pretty quickly for you. But for now this is a slower method to generate for free. Next up, they do have a Chinese website where you should be able to generate. I don't know if you can generate for free. I think you can, but the issue here is that you have to have a Chinese phone number and it would also help if you can read Chinese. Here's the sign in page. I Translated this with ChatGPT. You're supposed to put your phone number up here and then get a code. But obviously I can't get a code because I have an American phone number. If you have a Chinese cell phone, you should absolutely be able to get onto this site and generate with Wan 2.1. Our final free option here is Korea AI Limited free generations with a variety of video generation models, but they just added WAN 2.1. I've been waiting for this generation for a while. So it looks like their servers are swamped, but yeah, they definitely have Wan 2.1 on Korea now if you're interested in running this locally. Comfy UI did share a native support for WAN 2.1, which makes it really easy to set up and install the model locally. You can see they share a variety of locally generated outputs here that are all very impressive and exciting. This will all be linked down below, but this is a step by step guide on how to actually set it up and get it working in ComfyUI. Like I said, it's super easy. Download and install ComfyUI from the desktop app Download one of the WAN 2.1 models. Obviously for most of you this is going to end up being the smaller 1 3B model. Then you're going to download each one of these three models separately. They will literally just start downloading as soon as you click on them from this website and you're going to want to place them in their respective folders. If you installed the desktop app version of ComfyUI, this should be located in your Documents folder. You can then go ahead and download an example workflow, import it into ComfyUI and you should be ready to go. Here's the workflow actually set up in ComfyUI. I tried multiple workflows. I spent a lot of time trying to get WAN 2.1 working on my particular system. I've got an RTX 5090 yes, the brand new GPU. So Pytorch isn't officially supported fully for this GPU and we run into issues with this specific scenario. If I try to run the model locally, actually queue this up, you can see we end up getting a coup. Cuda error. I don't want to over complicate any of this, but by and large Pytorch does not fully support CUDA 12.8 on Windows just yet. Even with the nightly build. Yes I tried the nightly build. Not working for me just yet. However, if you are on Linux it should work with the 5090 and obviously a lot of those older cards, 40 series, 30 series should work without any issues. So yeah, I tried really hard to get this working locally. I tried other workflows, other methods. No cigar here, it's just issues with Pytorch and Cuda. GPU is too new. But like I said for the majority of you, the vast majority of you, you shouldn't run into issues like that. Some even are reporting to have this working on the RTX 2060 with only 6 gigabytes of VRAM, which is insane. So honestly try this on whatever GPU you have if you want to do this locally and let me know your results down in the comments. Anyways, let's actually do some quick testing and see how this model stacks up to Google VO2. Alright guys, so our first generation here is going to be in Korea AI, but Korea is just taking wa long, so I'm going to use a separate API for the rest of these generations. I would be using locally if Pytorch could support it, but at the moment we're struggling. Anyways, the prompt here a video of a man sitting at a table. On the table there's a huge pile of nails in parentheses the tool. He is casually grabbing handfuls of these nails and eating them quickly. Obviously an impressive feat here. And as you can see if you play the video, he's sort of just looking at the camera and casually munching on a bunch of nails from the pile. It's very close to what we described. You can see that Korea actually had to generate an input image here to be used to kick off our video. But yeah, it's a very clear video and I am happy and impressed with this result. Let's go ahead and now compare that to Google VO2, a paid, very expensive model that is only accessible on a couple of platforms right now. It's known to be the previous best video generator by the way, and we get a pretty similar video here. We've got a guy with a pile of nails and he's picking up some of the nails and munching on them. Although I will say in the VO example I've noticed that the nails are maybe a little bit more warped, a little bit more bendy as he shoves them into his mouth. I actually prefer the Generation personally from WAN 2.1 here and that's crazy considering VO2 is the previous best state of the art video generator and this one is free and open source. There's a little bit of weirdness going on with his shirt here. It looks like there's some glitter. I don't know if it's just because he's been eating too many nails, but. But yeah, at the end of the day the generations are close in quality, which is already crazy. But I do prefer this one. Alright, for the rest of the generations we're going to be using this WAN 720p API because I believe it's going to be faster. We do have an input image required here. The next prompt we're testing is Anime Ghibli style 2D animation of a giant lemon monster destroying skyscrapers for our image Input. Since it's required, I'm going to be using generations from Ideogram AI. It's my go to and they just released a new model, so I figured why not test it out. This is ideogram 2A and it did a pretty good job. We've got a nice Ghibli almost style anime monster. We'll upload this guy and click the run button. And here is our generation. That actually didn't take too much time, but we see the lemon monster move around. He uses a giant lemon to smash through a building. Although his hand does kind of like glitchily whip around his backside. But it's a pretty cool little animation nonetheless. I think this is very high quality considering again, open source model could run this locally on certain consumer GPUs. It's a great little demonstration. He destroys the building like we ask in the prompt. I am impressed. Even more shocking, take a look at the Google VO2 generation. Yeah, we've got an anime Ghibli style lemon monster. We've got some buildings here, but he kind of just like puffs up and gets angry. He doesn't actually move or destroy any buildings. This generation is so much better in my mind. It is a lot more animated. It's a lot more detailed. There's more going on. I am a big fan. So is Wan 2.12 for two right now? Yes, it is. Our next input image. Very simple prompt. Close up video of an orange tabby cat eating a lemon. By the way, guys, we could also mess with all kinds of settings because it is open source. So you might even be able to improve these generations beyond what you're seeing today. That is one of the many beauties of open source things that you can especially run for free on your own hardware. And our next generation is complete. Nice close up video of a cat munching on a lemon. It looks a little bit kind of like stop motion or claymation because the frame rate is a little bit lower than you'll find with most video generators. That could be changed or increased with the settings, but for the sake of time we'll keep it at 16. But it's pretty decent video overall. He's kind of biting into the lemon. It's a little bit maybe unrealistic in some ways, but you can actually see on the crest of the edge of the lemon that the hole inside of it does expand as the cat munches down, which gives the effect or implies that he is indeed eating the lemon. And it kind of looks like how A cat normally eats, but definitely not as good as the VO2 example. Here's the Google VO2 generation. Yeah, I think this one is definitely a little bit better here. You can see the cat is just sort of licking and munching away. It looks a lot more realistic, I think, to how a cat actually eats. So I'm gonna give the win here to Google VO2. I think this is a more realistic generation. I think it looks a little bit better to be real, though. Stonkingly close, though. I mean, WAN 2.1 comes very close to the leader, cutting edge VO2. So at the end of the day, I would give it closer to a tie, considering this is open source. Pretty incredible though. Not gonna lie. Next up, we have found footage. Cctv, gas station clip, man riding atop a giant bullfrog. And this will be our input image. Yeah, it doesn't really look like CCTV footage. I know. Neither does the Google VO2 generation. We'll go ahead and send this one on through. Alrighty. Man rides a bullfrog at a gas station. You can see that this one definitely was a struggle for the WAN 2.1 model. Absolutely a flop on this generation. We could always run it again and see if it does it better. But yeah, this is definitely not looking too hot. We can see the gas station warp and move crazily in the background. The frog is kind of moving along nicely. The man is pretty still overall, but definitely not a good generation for WAN 2.1. Undeniably, the win is going to go to Google VO2 here. This is a pretty great video with the guy riding the bullfrog. It's pretty realistic overall. Obviously none of these were CCTV footage as I described earlier, but yeah, he's definitely riding the bullfrog here in VO2, and I think it looks a lot better. Not nearly as much morphing or detail issues. Right, next up, we have a cinematic wedding ceremony. Shrek and Bigfoot get married. Very silly prompt, but this will at least test copyrighted characters and mythological legends. Go ahead and send that on through. All right, we've got Bigfoot and Shrek getting married here. I actually think that the movements of Shrek as he sort of talks and moves around are pretty reminiscent of the character from the animated movies, which is cool. And Bigfoot looks pretty stable too. I like the twinkling lights here. It's a beautiful scene. Of course, I think it is a fair argument to say in this specific case, Ideogram is doing some heavy lifting here, defining Shrek and Bigfoot as separate characters where VO2, as you'll see in a second, kind of struggles with that. But yeah, overall this is a pretty impressive generation. The background still, the characters are moving. I even like the slight sway on Shrek's clothing. I'm happy. Here is the Google VO2 generation. We can see that there is something cool at least with this generation going on. We see the swooping of all of the plants sort of growing across this circle thing. We actually see the characters walking and the camera moving. It does look like more of a cinematic wedding ceremony, I'll give VO2 that. But obviously we are having some character and subject confusion here. Now if we could have uploaded that same Ideogram video, we'd probably get a better result with VO2, but we don't know for sure. So Ideogram might have been doing some heavy lifting for WAN in this circumstance. But yeah, Honestly, I think WAN 2.1 still takes the cake. Silly prompt regardless. But our next prompt is going to be a home VHS video Christmas opening a present to reveal a mini nuclear reactor inside of the present, which is glowing green and radiating. So the perfect Christmas gift for any child. Let's go and send that one on through. And here we are with our WAN 2.1 generation. I think this came out pretty stellar. It opens in a little bit of a weird way. It's not tearing off wrapping paper like you'd expect, but there's definitely some sort of glowing green reactor esque device on the inside and it's spewing radiation out on Christmas morning. So I think that it did a pretty stellar job with this. You can see though that the box does kind of clip through the tree a little bit as she opens it. But I mean overall, for open source, for being free, this is amazing. I'm really impressed. Especially keeping that child stable and keeping the hand stable as she opens the present. WAN 2.1 is really a game changer. Man, this is pretty eye opening. Now here is what we got from VO2 and I think you'll see some issues with VO2 in the fact that it's trained on a lot and I mean a lot of YouTube video data. So, you know, this isn't the typical Christmas morning that we would like or expect to see. It's actually like a YouTube demonstration of a nuclear reactor mini nuclear reactor product. And I think the nuclear reactor looks a lot better, sure in VO2, but this definitely isn't Christmas morning. Now again, there is the argument to be made that Ideogram's initial input image is doing Heavy lifting here, but I couldn't even choose an Input image for VO2 in this circumstance if I wanted to. So I'm just gonna have to hand the win on over to WAN 2.1. This is not Christmas morning. Yeah, I like the mini nuclear reactor. I like that. It's like a guy reviewing it. That's a cool video. And it looks very real, but it did not follow the prompt as well. Next up, this is going to be our final one. Cinematic movie footage tracking a woman with an umbrella in New York City. She looks mildly displeased as lemons are plummeting to the ground from the sky, bouncing off her umbrella. Difficult prompt. Let's see if WAN 2.1 can handle it. And check out the generation. I think this came out pretty decent. We've got the disgruntled woman right in the center. We've got the lemons falling from the sky as it rains. It seems to be not only raining lemons, but also actually raining. There's little raindrops. We can even see some cars moving around in the background. But they're kind of growing and morphing a little bit. They don't quite look right. You can see if I actually move my head here. Yeah, there's just some. A little bit of weirdness with the cars down there. But overall the main goal here, the lemons falling and bouncing off. It's close. It's close. I would love to see better physics here. I would love to see the lemons really like pounce in and bounce off of the umbrella in a realistic way. This does look like something more edited. The lemons look definitely a little bit more cut over the original video. Let me know what you guys think. But here is the VO generation. I think that this one is a little bit worse. The lemons actually look more fake in this demonstration. I think the background looks more real. Maybe the woman looks a little bit more real. But there's lemons aren't even touching the umbrella in this. And they definitely just look like a plastered over animation of some form or function. They don't look very real. I mean, to be fair, neither does this one. But at least this one has more lemons and they're actually somewhat touching the umbrel umbrella in some of the cases. It's a difficult prompt for sure. Neither of the generators really come close to nailing this one. But WAN 2.1 is close enough to VO2. I think that a lot of people might actually prefer this generation. So yeah, at the end of the day, guys, WAN 2.1 is an absolute butt kicker of a model. Open source runs on local hardware, can be used for free on certain websites. It's definitely cheaper, more efficient to run than a lot of closed source and other open source video generators. Just mind blown this absolutely could be the stable diffusion moment for video generation. No doubt about it, the community is already hard at work with this model, trying to get things like Laura's to work correctly with it and hopefully optimizing it even further. I'd also love to see fine tunes for this specific model as well. I'd love to see like an anime fine tune. I bet that would be really cool and actually work quite well because this model seems to lend itself pretty well to animation in general, at least from the examples we saw today. But yeah, get excited guys. This is nothing but good news in my book from Wan 2.1 here. Hopefully I can get it working locally on my machine very soon. I'd love to test it out locally and include it in an upcoming video. In other news, GPT 4.5 also drops today, so expect a video on that pretty soon. I think that might have to take precedence over my son at 3.7 video. Thanks so much for watching guys. Again, let me know about your results running this thing locally on your own machine and I'll see you in the next video. Goodbye.