Open Source AI Video BEAST! Magi -1 Autoregressive AI Video Gen

A brand new, absurdly capable looking open source AI video generator just dropped. When you guys see the first couple of examples this thing has to offer, I think your minds are going to be blown. This is actually the tweet that brought this to my attention. Sand AI just dropped MAGI one, a groundbreaking open source video generation model. They claim unmatched control over timing, motion and dynamics. And take a look at that first generation. It's about seven seconds long, but man, it looks real. We've got this young. She kind of puts her head down onto this plant and the plant reacts in such a realistic way, the physics kind of jiggles around. It doesn't look like slow motion physics, it doesn't look like it's from a video game or something. That looks like real footage. And not to mention when the girl looks back up at the camera, look at that. The facial expression is spot on. It looks like a real video of a person. These eyes, they look like they have soul behind them, dare I say. And the way the shadows from the tree that she's presumably under cast on her face make this even more convincing. I mean, this is a staggering first generation to lead off with, but let me tell you, this clip is far from the most impressive thing you're going to see today. Let's take a look at the rest of the thread. MAGI also claims infinite extension. No more fragments or stitching. It's apparently the only model that can offer infinite video extension, seamless full length storytelling. Right, And I kind of disagree with that. Plenty of video models, including open Source ones like WAN 2.1, have input and endput video frames, meaning you can take the last frame of whatever video you just generated and use it as a first new frame for a brand new generation. I believe the reason MAGI is claiming this infinite video extension and claiming they're the first to the table with this is because the architecture under the hood of this model allows it to do this extension a little bit more natively. It doesn't need to rely on necessarily the exact same frame to continue a video. It can rely on the context of the rest of the video as well, giving you a better extension. So while it's not technically the first to offer this, it might be the first to do it this well. They're also claiming precision timeline control. The only model with second level timeline control, meaning you can craft every moment exactly as you envision. This is a little vague. I'd love to know exactly what they mean by this and we are going to find out today. The example that Angry Tom provides for this. Again, another stonkingly impressive generation. We've got this giant sci fi eyeball in this sci fi city and the way it moves and kind of blinks and shutters is just creepy. There's plenty of detail in this eyeball, especially in the way it blinks and moves. It's not something that I would expect to be in the training data at all. Impressive. Lots and lots of detail, great movement and perfectly still background. And don't forget guys, this sucker is fully open source again. Another thing you're going to notice about the model that Angry Tom points out here is that you're not seeing a lot of that slo mo esque footage or limited dynamics. They are claiming precise prompt execution and industry leading video generation performance. And I gotta say guys, taking a look at this example of the girl playing the violin, it's stonkingly impressive. You can see she's kind of shimmying her fingers. It's moving quickly. Yet other parts of the video are moving nice and slow. Definitely not slow motion video. This is what it looks like like when someone really does play the violin. You will notice here though, a little bit of movement on the strings. Not too much clarity or coherency overall in the rest of the violin. There is no way that this generation is going to be as detailed as a real, you know, 4K camera recording of someone actually playing the violin. But yeah, a difficult task, a difficult video generation task. And this open source model really nails it again though, cherry picking. Cherry picking. Right. We gotta test these models for ourself. Don't worry, hands on is coming up. So yeah, Sand AI's Magi One is fully open source and that means it is full model weights open and full inference code open. You can download both for free. And honestly with something of this quality and caliber level, being open source is just a cherry on top. A lot of companies would keep this one closed source and try to take down cling AI for example, with their own platform. And they do have their own platform. But hey, it's also open source. It's really a two for one. All right, let's learn a little bit about the model under the hood. Magi 1 generates videos autoregressively. It predicts a sequence of video chunks. And these video chunks are basically fixed length segments of consecutive frames. You know, a video is made of frames, basically just a bunch of pictures playing in sequence. It's trained to denoise per chunk noise that increases monotonically over time. Something that changes in A consistent unwavering way, either continuously increasing or continuously decreasing. Okay, stay with me. Magi 1 enables casual temporal modeling and naturally supports streaming generation. Strong performance on image to video tasks, high temporal consistency and scalability. And they claim it's not just up to the autoregressive training here. To achieve what they did here, they needed several algorithmic innovations and a dedicated infrastructure stack. The further controllability is made possible by chunk wise prompting so you can get smoother scene transitions, long horizon synthesis in fine grained text driven control. Magi1, they believe offers a promising direction for unifying high fidelity video generation with flexible instruction control and real time deployment. Man, what an introduction to this model. And it's open source. They do have another really good demo video which I also want to show you guys. But let's wrap up talking about the technicals. So it uses a transformer based vae, but the secret sauce that everything is built on really seems to be that autoregressive denoising algorithm. So you can kind of get an idea of how it works based off of this. Like we said, it generates chunk by chunk. Each one of these sets of images is a chunk, right? Once one chunk reaches a certain level of denoising, the next one begins. The pipeline design enables concurrent processing of up to 4 chunks for efficient video generation. And of course understanding what's going to come with the future chunks before the previous ones are done, giving it a better understanding of making that motion over time. Don't know if I'm explaining that super well, but that's, that's what I'm kind of picking up on now. Sand AI did run their own in house human evaluation on the MAGI1 model, comparing it to the heavyweight open source champs huanyan video wan 2.1 hail you cling. Take a look at this. The first hail you one, we've got a win rate of about 30% overall going to magi1. Hail you only winning 22% of the time. With 48% of the time, the results were comparable, about the same in terms of overall quality. But it doesn't matter which specific benchmark mark you pick. Motion quality, instruction following visual quality, MAGI one technically wins. There's a pretty big spread here though, showing off that a lot of people still think that Hyl you is nearly as good as Magi1 and in some small cases better. But for Huan Yan, they don't get off that easy. Overall, we're looking at a 44 win rate for Magi 1. 45. The same and only a 12 win for Huan Yan Video. And the rest of the benchmarks are even crazier. 14% versus 2%. 15% versus 1%. Surprisingly, the only model that's been really able to kind of keep up with Magi one is cling. And a lot of people swear by cling. It makes sense. This is just cling 1.6 also not the new Kling 2.0. Don't have any evaluations on that. But cling is able to win overall at 27% versus 17% win 2.1, which is what I presume to be the best, highest quality open source video model until today. Just barely gets beat out by MAGI1 on these benchmarks. Alright, so now that we've talked about benchmarks and how it works under the hood, let's take a look at their professional demo video. This is a little bit more edited, a little bit more put together. Kind of shows off what you can get up to with this open source model.

Hey, you know what? I've been a great actor since I was born. I don't need a break. I'm not afraid of danger. I work hard. I don't complain though. Before every performance I had to take your boring auditions. Crying, laughing, crying and then laughing, Fighting, dancing, flying over the place Sometimes a hero, sometimes a fool. Dreaming a world will make believe. In this place to live or to die Question your pondering My destiny can be pinpointed to one second or do a long termism Choose a path Keep going, keep going and fly continuing in the ever generating future we collide with each other pushing the boundaries of imagination and your preference for each choice generates the world this new world it's called magic.

Pretty great little demo video. Honestly, I'm impressed. They're showing off a consistent character here. This one actor. Right? I think this is a pretty funny scene as well. But you know, he's AI, he's tattooed with it, he's back in time, holding a sword in a battle, he's dancing around in a flower outfit, fighting dragons, catching bullets, all kinds of stuff. So as I mentioned, they do have a website where you can actually generate and use an interface with their models. But of course they are open source, so what about running them locally? Well, here are the models we have to work with. Starting at the bottom. Magi 14 5B. This is their smaller model. This can run on one singular RTX 4090. Not too bad. Still, that's a hefty amount of VRAM, right? 24 gigabytes. But consumer GPUs can run a variant of this model, although it's a smaller one. Hopefully though the open source community will get their hands on this and tear it apart to make it run on virtually any system. There are a ton of frameworks and a ton of different options and the open source community has done a great job at getting open source video in the hands of folks with consumer regular computers. Anyways, we then move up to the 24B distill and FB8 quant. This is requiring quite a bit more. Right. We're looking into server farm territory. H100 or H8 hundreds. You're going to need at least four of those to run this quant or eight RTX 4090s. Totally crazy. And it just gets more insane from there. We've got the 24B distill model. Again you're going to need eight H1 hundreds or H8 hundreds and with that amount of compute you should be able to run any of the other ones as well. So if you want to run this thing locally today, you either need a monster consumer PC, the minimum being an RTX 4090 or you're going to have to wait a little bit for the community to distill these models further, do some enhancements on them and get them running on way less vram. If I see anything cool though, I'll definitely let you guys know in an upcoming video. All right, let's create with Magi. We've got a free plan right off the bat giving us 500 coins. New project. Let's go ahead and toss me in there. Yay. It seems like they're using a canvas style board. All right, we'll click a plus on there. Image to video duration time Variations 1. How many variations will they let me do? 10 for 300 credits. Okay, I like that. They'll just let you do whatever you want. But let's keep the variation at one for now. I don't want to burn through all my credits. It looks like it's going to cost about 10 credits per second. Alright, let's bump this up to 5 seconds. A lemon falls from the ceiling and the man catches it in his mouth. Lemon juice flies everywhere. What happens if I click the plus over here on the side? We can do an extension of this video. I see how this works. Very interesting canvas style board way of working with AI videos. Not sure how I feel about it just yet. It looks looks pretty clean right now. All right. Our generation estimated wait time is about four minutes right now. Hey, I'M glad that they have some sort of a free plan. At least a lot of companies wouldn't give you any free credits at all. All right, let's get another image. And this time one generated by GPT4OMNI. Yes, that is me and Snoop Dogg. Did I say GPT4O generated this? No, it's actually real. Me and Snoop Dogg are just friends like that. All right, we'll do this one only at three seconds. Just a quick little test. Move this down here. It looks like our first one is done. Oh, lemon falling into my mouth. Kind of almost. Oh, you can see the enhanced prompt as well. It uses AI to enhance them. You gotta think sometimes whether all of this is worth including in the prompt. I mean, it is their model and they would know best, but I feel like this could kind of overload things a little bit and prevent some of the real motion you want to happen. But yeah, five seconds, I open my mouth. I'm pretty consistent overall, but the lemon is kind of just dangling and front of my face. And then there's some confetti raining down yet. Like, you can see that this is hanging by a string. And in the prompt it describes exactly like my background and stuff like that, which it really doesn't need to do. I might start disabling enhanced prompt, at least for now. All right, this generated very quickly. How does this look? Yeah, me and Snoop are just hanging out. A little bit of weirdness with the smoke, I would say. But for three seconds, that's not too bad. Alright, let's move this way. We're gonna take this image of the rear end of a car. Let's do a longer generation. 10 seconds. Can I do like 120 seconds? Is that even possible? No, it looks like the max is 10 seconds and then you have to extend them, but the extensions are more native. Again, think back to those chunks working simultaneously. We'll keep it at 10 seconds. I'm actually gonna use the same prompt I used when testing Gen 4. The camera moves towards the left hand side of the car, then stops facing the rear of the car. The car speed, leaving smoke and a fire trail behind the tires. In cinematic fashion. Let's turn enhanced prompt off for this. I want to see it just raw. It seems like they do have a high quality versus a lower quality mode as well. If you want to try to save on some credits, we're gonna keep this high quality. Right now though, I want to see the best of what this model can do. I like that you can queue up as much as you want as well. Let's grab another Image generated by gpt4o, the Armless Man. Let's see if it can keep him armless as he runs away into the distance. This is something I remember gen 4 not being able to handle. Or pretty much any AI video generator again. Turn, enhance prompt off. Man stares into the camera for a beat, then sprints away in the distance. As he sprints away, the camera starts to pull away and fly upwards, showing the landscape. Send that off through. Alright. And let's try something creepy as well. Yeah, that is another GPT4O image. That's me sitting in a room filled with lemons. Very creepy. Creepy VHS footage. Man buries himself in lemons like a crab. Keep this at five seconds, generate. Let's also see how this thing does on animation. I've got this little robot riding around in a rocket. Boom. We're gonna do another longer generation of 10 seconds. We must be running out of credits by now as well. Hey, they were free credits though. This was a prompt that I remember. Gen 4 Absolutely. Nailing. Camera follows a cute robot riding a rocket ship through space. The rocket lands on the moon and the robot gets out of the rocket. Cute 3D animation style. And we'll send that through. Alright, we got the man staring at the camera for a beat and then sprinting away into the distance. So he stares at the camera for a beat for sure. And then he begins like he wants to sprint away. Oh, he does sprint away. And then it tries to start to focus on the landscape. This was only a second generation, so we don't know if it would start to fly and pull up showing off the landscape. But the man sprints away and waits for a beat. He does not sprint into the distance though. Still, the quality kept him armless. I gotta say, that is impressive. Gen 4 immediately just morphed his arm into like a full working arm where this can actually keep it separated. That is mighty impressive. Alright, the car generation is also done. How's this gonna turn out? Okay, well, it appears that it's struggling a little bit, keeping the car consistent, but once it peels off, that's pretty fun. I know. Maybe cars aren't this thing's strong suit. This is concerning to me the way that the tire and the ground is shifting. It seems like the model can't tell whether or not the car is mo moving or not. Well, that's gonna be an issue here. The car is not moving, folks. But the model thinks that it could be based on its training data. So that's something that they might want to look into improving. Fine tune could probably fix this up pretty well though. But yeah, just to play it again, you can see little bit of glitchiness. Once it tears away though, it does leave the fire in the smoke trail. And yeah, that's pretty fun. I do like that after it starts to move. Alright now, my creepy generation. How are we looking here? Oh God, this is really creepy vhs. Oh my God. Why am I just morphing into the lemons? Oh, that is just scary. I. It is very creepy. So I'll give it that creepy VHS footage for sure. I don't know if I'm burying myself in the lemons like a crab. That's something I would complain about potentially. So we can of course extend these videos though. So how about we extend this one of the guy? Alright, let's try to continue this to see if it can pull the camera up and show off the landscape. Ooh, and our rocket animation is done. Let's see how this thing tackles 3D animation. We start off here. Can see that he is flying towards another planet. It's supposed to be the moon. Oh my God, he is rotating and morphing. It's like a transformer. Okay, well that was not a great generation. Interesting. Okay, you can see how the background kind of morphs over time too. Immediately again as the motion moves. I wonder if more time in the oven would fix this. But there it goes, trying to go on the moon and it just sort of flips out and goes crazy. Glitch mode. The AI doesn't know what to make of that. Maybe 3D animation is not this thing's strong point. We are dealing with new architecture here, so I'm not surprised that it's not behaving exactly like the traditional AI video generation models we're used to. There is a strong likelihood that this model might need to be prompted a little bit differently than we're traditionally used to. Okay, and then this video extension, no, it does not fly up and show off the landscape. Instead it kind of just falls down slightly. And keep in mind as we're generating with this, there's no real way to tell which one of the models they're actually running under the hood. And it could be different than what we actually saw in the demo videos. Alright, well, I've burnt through almost all of my credits. Let's observe the plans. First of all, they have two options. You can either go with a sub or you can buy the credits directly. I like that they give you the option they do give you 20% off if you pay yearly, but I don't recommend that at all. With the rate of technological innovations in the AI video generation space, I wouldn't pay for any yearly plan, not unless I had a really big project underway. So looking at the monthly ones, the cheapest one is going to be 10 bucks a month for plus. That gives you 1500 credits a month, three times what you get with the free plan. And apparently these credits do renew, which is also really nice to see. That's about 2.5 minutes of video per month for 10 bucks. Actually, as far as AI video generation websites go, that is pretty darn cheap. If you move up to Pro, that's 9,000 credits a month. Obviously, the more you spend, the better a deal on the credits you get. This is about 15 minutes of video generation per month for the PRO plan. And again, this is calculating on the high quality video generation. So about 10 credits per second of video generation now. Premium, 100 bucks per month, 20,000 credits. You're looking at 33 minutes per month of generated video. I gotta say, I could see someone maybe going with the Pro plan if their video generations work very well with this particular generator. But if I really want to pay for something, I'd probably start off with this $10 per month plan. It's pretty darn cheap and you're still getting a decent amount of video generation. Now let's check out the credits, because you can just buy the credits outright. Yeah, you're definitely getting a worse deal if you buy credits outright. A thousand dollars to buy 200,000 credits. Who's emptying their wallet? Yeah, 10 seconds of video is about a dollar actually cheaper than VO2 still, which is pretty wild. I like that. It does give us a pretty solid idea of how long these credits are good for. Two years at least. They won't disappear in a month. Now on this website, they do have an assets section which seems to be powered by the community. They've got a bunch of featured assets that you can use to start any creations you want. Just help you ideate. Of course, your uploaded media is also saved in here as well as all of your different generations. So is MAGI one the new open source king of AI video generation? Well, kind of yes and kind of no. The model is capable of phenomenal outputs, very lifelike, high fidelity, good understanding, and you can continue infinitely. But in my personal testing, it doesn't seem to be as easy to manipulate and get a good result out of compared to what we're traditionally used to with AI video generation models. It's auto regressive. It is a little bit different under the hood and that means that we're gonna have to learn how to maneuver this thing properly. I do love that they have a completely free trial on their website. I recommend you guys hit that up. Start generating and share your generations to my Discord server so we can see how good this thing truly is. The whole community needs to get together for that. Comparing directly with Gen4, I think Gen4 is still better. Overall. It's a lot easier to use in terms of being able to just send it out with a prompt and get a good result back. This thing seems to be more like an untamed wild beast. Capable and very potent but difficult to control. At least if you don't understand it. Hey, it is fully open source. We got weights and inference code. Couldn't ask for more than that. So good on sand AI for doing that. If I had to pick a favorite AI video generation model, one that I would want to use today locally, I'd probably still stick with WAN 2.1 just because it has more experience with the community. There are workflows, there are new things that exist for WAN 2.1. This is a little bit too new to have all of that open source community support just yet, but hey, it just came out yesterday so. So we'll wait and see. Thanks so much for watching folks. I hope that this video was useful to you guys in some way. Let me know what you think and I'll see you in the next video. Thanks for watching. Goodbye.