Happy Horse 1.0: нова опенсорсна модель відеогенерації перемагає Seedance 2.0

Hey guys, welcome back to the Matt Vid Pro channel. A new video generation model has appeared on the artificial analysis leaderboard and as you can see, it's got a lovely name, Happy Horse. This is one Happy Horse, guys. Taking the number one spot over Dreamina Seance 2.0. A very exclusive and high quality model, which now is available in the US. By the way, guys, we're going to talk a little bit more about it later, but Sea Dance 2 is finally coming to the United States. Regardless, yeah, this happy horse has a higher ELO score than Seance 2.0, and it isn't by some marginal amount. You can see compared to Cling 3 and Sky Reels, Seance only edges it out by a little bit, but right now, Happy Horse has almost 100 ELO points over Seance 2.0, and that's with a lot of samples actually. So pretty surprising. Although I have heard people claim, oh, you know, people are just trying to vote for Happy Horse on purpose. Although I think it is genuinely a very good model. I've seen it lose in a certain couple of areas though. Who is Happy Horse and why is this horse so damn happy? Well, it actually has officially been revealed as of this morning, but Brent Lynch has got a couple of nice examples from the video arena. Brent correctly guessed that this was an Asian model, but is wondering if it's WAN 2.7. Not quite. Is it a Cance 2.0 killer? No. And I would agree with that. As Brent points out, it is more in line with the recent Cling models. I suppose in some ways, but I actually think it's better than the recent Cling models. It also does have native audio, and it's great audio at that. Let's roll through some clips. This model, by the way, does appear to be 720p. And you can see it's got great detail, great resolution. The imagery that comes out of this model and the movement, I think you're going to be really impressed with multi-shot. Definitely a strength you're going to see here. I love how still this first shot is. He walks right up, but it feels super believable. Very solid background, the cut to the puddle, and then the step, accurate boots, pretty realistic crunch. I think the sound is decent. It's a little cartoony, a little movie mode, but definitely not terrible. Left a little bit of a weird blue trail. I don't know if that is just Twitter compression or the model itself. I assume this is honestly Twitter compression. And back to the crushed piece of ice. slow motion. Not bad. Okay, the the voices at the end were a little dramatic. They're like screaming as if meteors are falling out of the sky. A couple of things I'm noticing. This is a more difficult test than you actually realize because the physics getting the speed right for the basketball. I mean, you can throw a basketball at any number of speeds, but I think watching this, this looks like it's actually slow motion footage to my eyes. That's the the criticism here. But again, everything else, so much detail. We've got them, six people. After the ball hits, we still have those six same exact people. This looks a little bit less like slow motion. I think the net physics here are really impressive, but not 100% perfect to what would occur in real life. I think they come a little bit too close towards the viewer instead of slamming back more. But I think this part where the ball comes through really good. Okay. Supposed to be kind of science techy. You can tell some of the audio like it doesn't know what it wants to do with it almost. Seance 2.0 has better audio. I'm pretty sure that the entire telescope isn't correct for a real telescope of this size or whatever. And I'm pretty sure she's supposed to look through this. I'm not sure if this should even be there in the first place. But there's no morphing. The human looks very believable. 2D animation, guys. Look at the almost crayonesesque drawing of this scene. It held up from that initial image input very well. firepowers. That looks pretty good. A couple of strange frames here, but like shooting the fireball, like watching it, it's good. This part as well, where the flame is floating, you'll notice with this model on fast motion, there's a lot of warping and mushiness, but it corrects itself very quickly. And cases like this, you kind of barely notice it. Honestly, not too much to say with this one. The robot arms moving look okay, although they are pretty inconsistent. Like, I don't know what's going on here with this arm. Yeah, like we said, not a seance 2.0 killer, but I still think this is very much impressive and definitely a big step forward. Now, why do I say this is a big step forward for AI video generation? Well, because ladies and gents, we now know what this model is. And better yet, we also know it's going to be released open source. Yes, Golf Collab. Very excited. Higher ELO than Sea Dance 2.0 on the leaderboards. So, statistically and artificial analysis, it is better than Sea Dance. To me personally though, like my take a little bit worse than Seance, but open source, I mean Sance is closed source. We can't even get an API for Cance. We just now got access in the US to Sea Dance. And now Happy Horse is coming out open source. Man, like I said, this horse is freaking happy. It is basically open-source seance 2.0. It's got the resolution. It's got the sound. Maybe not as perfect, not as good, but it's 80% there. I think Happy Horse is a very memorable name. Like Nano Banana 2. They should just keep it. But who is behind it? Alibaba or Cowba. All right, let's mess around with Happy Horse here in the arena. You don't know exactly which model you are going to get at any given time. This is free to use as you are just ranking and voting the models, but you can't put any prompts in yourself. We are on no audio mode. Eiffel Tower time lapse. Sun is setting. City comes alive at night. This one is more realistic feeling. I feel like it's just closer to the prompt. All right. I prefer this one. Oh, okay. That was VO3.1 light, that one. Now we got penguins toboggoning. Some light AI mushiness over here on the left, but they're not really doing toboggoning. This one, at least they're kind of getting on their bellies. All right. Yeah, we got that in that first opening scene, but then it quickly just diverges into other stuff. They're both pretty terrible. You not going to get the happy horse every single time. I mean, just looking at the shadows alone almost, I feel like you can tell that this one is a little bit better. This is still pretty decent, though, honestly. All right, this is supposed to be changing through the seasons. While this one has worse image quality, it actually followed the prompt. Okay, so running the finger on the carved chair leg, walking through the boards. I think this one's trying to do a little too much. This one is more accurate to the prompt. Oh, this is definitely one of the the good prompts that I saw. This guy turns around midway and this one actually completes the whole ring. So, we're going to go with this one. And that was actually Happy Horse. Okay, it destroyed VO 3.1 fast preview VR training simulation. This is a cool prompt. I like that this is actually in first person as well for this one. So, we're going to go with this. If you want a better chance at getting the Happy Horse model, switch to with audio as not as many models have access to audio. So, Happy Horse is one of a few. Okay. Some gross liquid. This one definitely has better audio. Yeah, we're going to go with this one. Yeah, that was Happy Horse. Yeah. Okay. That's pretty terrible for the dominoes falling. Oh, this is almost just as bad. Okay. Um, this one's probably going to be better. That was Happy Horse. See that? Sometimes Happy Horse is like just terrible. I'm like, dude, that was a terrible generation. So, yeah, hopefully Happy Horse is releasing publicly very soon. We could just mess around with it on the arena, check out some of those early samples. It looks like a great model, and I'm really excited, praying that it's actually going to come out open source. A fully open source release. I mean, that would be very, very sweet and that would make all us horses happy. So, like I said earlier, officially Dreamina Seance 2.0 is rolling out in the US. Finally, they waited and made the US the last final country, probably because of the copyright laws, I assume. Uh, but they've got some serious safety filters on Sea Dance. They give they're giving APIs out to various companies. I've been using it a lot on Polo AI. It's coming to Runway ML, which is super exciting. That's one of my favorite websites. This will be the first time you can really get a true free trial, free access to Cance 2.0. And this is across all platforms, the Cap Cut app, desktop, and the web. They're running a pretty sweet promotion. 90% off the first month of the pro plan. So, you get like a a free month basically. Seed dance 2.0 is a very very nice model. It feels like a more raw version of Sora 2 almost in some ways, but like I said, they really ramped up on the safety features, the censorship, certain violence, certainly gore, definitely nudity, and of course, no uploading of real human faces, not even your own. The sketch method seems to kind of have already been patched a little bit. So, it's like you have to take this human face and then severely edit it down and then upload it in and hope that the model can rebuild the face from scratch. It's a pretty big pain. Regular text to video though allows you to just generate everything you want. But it's of of course the issue is character consistency. So, animated characters, it's easier to get away with right now. But yeah, the safety filters are are pretty insane. Like I said, I've been using this thing on Polo AI, guys. This right here is Cance 2.0 fast. Yeah, there's a faster model that's like half the cost and maybe 70% as good. Something like that, little character concept. It's pretty great. Straightens his tie up. Here is Cance 2 fast again. Not even the full big bad Sea Dance 2. I'm just not really in the amphibious mood today. >> It's very good. It It's great at that movie cinema type of feel, line delivery, and that music. One day we'll be able to string these much easier together, hopefully with less restrictions, I guess, or just be able to do much longer natural generations. 15 seconds is the limit. One day that limit will be, you know, 5 minutes and then 10 minutes. Eventually it will be hours. That's far down the line though because AI video takes a ton of compute. All right, so I want to cap this video off with GPT image 2. This is not out to the public yet, but it's currently being A&B tested inside of Chat GPT. I ran a bunch of image generations and I didn't get any A and B testing, so I haven't been able to make any images for myself yet. But this GPT image 2 is looking like a mighty sweet image generation model. Possibly stronger than Nano Banana 2, Nano Banana Pro. It might be seriously good. Hopefully they don't labize it or neuter it before release. Insanely strong outputs, guys. Really excited to get access to this one. Obviously, we're looking at a Minecraft screenshot, but we're in Claude headquarters. So, internal document Claude Opus 5. We've got text generation, but it's also in the correct Minecraft font and almost pixel perfect, right? Like all this pixel art that makes up the the text characters is like perfect. Next generation model, advanced reasoning, blah blah. Confidential stack of papers here, the geometry on this is a little less believable. You can see down here, Claude has joined the game. That is accurate. Again, perfect font and color. Pot bar down here looking pretty much exact. It's a little different in some subtle ways, but pretty close. Correct lanterns. Some of the blocks are recognizable. But yeah, the most impressive to me is the clawed headquarters with the accurate pixelated logo transcribed onto a Minecraft wall. Again, almost pixel perfect here. This thing could be very strong with actual pixel art. Here is another one. Windows screenshot. Like very flawless in a lot of ways. the logos, the icons, the font. It all looks almost indistinguishable from an actual Windows screenshot. There are subtle giveaways, like the text being too close to a side of a wall or just some like mushiness down here, but it's very close to being perfect. And look at how much text they're able to stuff in here and have it actually all be correct and legible. video memory, graphics controller, remote desktop controller, IDE, Ubuntu, you know, showing the virtual box manager here, the VM workstation. Very, very cool and hard to believe in some cases all AI generated and not a real screenshot. This one is a little bit less convincing, but I think it's pretty obvious this is a Fortnite screenshot. We've got Elon Musk, Daario, and Sam Altman playing trios here in season 2, Chapter 5. Yeah, the amount of correct text here again shocking. Absolutely astonishing. But the battle bus looking good. You know that piñata animal and then you know they actually do represent Sam Alultman, Daario and Elon. That is pretty decent likenesses. However, they don't look like true Fortnite characters. They look like photoshopped amalgamations, some kind of meme wear. And the reason that this is is because the model probably is looking at what it's actually been trained on, noticing like, hey, this type of concept is typically this, you know, meme stuff that's imperfect. I don't really need to go all the way all out. Maybe the prompt just wasn't detailed enough. This is no doubt a reasoning during or at least before the generation of the image similar to Nano Banana really on a whole different level. I'm really shocked by how good it is at UI and text and just creating a cohesively whole image with this much going on. Yep. And then here's the classic Steam store page. Almost indistinguishable. Halflife 3 is out now to to wish list, you know, coming soon. But it's like crazy how accurate the Steam UI is. Is this game relevant to you with all of this stuff basically flawless? Different icons showing all the different things that it supports. the actual events and announcements coming soon. Add to your wish list, developer, Valve, like they haven't really changed this interface in years and that's why this is coming out so so crisp. But yeah, an impressive model because I've never seen screenshots this good from Nano Banana Pro or Nano Banana 2. This might very well be the best image generation model we've ever seen in terms of raw ability. Mark Crushman brings a few more samples. Superb text rendering and realistic photo look. It's much more convincing and it doesn't have that yellowy filter or that AI slop overdone, overexaggerated feel like we've got a Medicare request and it's doing this flawlessly. Like that's pretty crazy. It looks very real. The handwriting is almost too perfect to believe. Everything about it feels a little too set up and organized, but it's very convincing in terms of lighting and detail and like, wow, look at the text. Look at how good that is on the bottom. Oh my gosh. Medicare in Australia. Yeah. Pretty insane. Yeah. Or this this borehole log. Like are we kidding, guys? This just looks like a document. Like just 6 months ago you showed me this, I would have been like, that's not AI generated. That's a real document. Look at that. You tell me that's not a real doc, bro. Pretty crazy. It is so crazy. Openai has seriously cooked something up for sure. some Ka logs, corn flakes, only a little bit of mushiness at the top, but yeah, a raw feel. It's not overdone in terms of saturation, not too much processing. Realistic. And then we've got Obama and Trump, you know, they're they're hanging out. Very realistic feeling. So, obviously, this demonstration showing a little bit more of the safety side of things, what this can generate realistically and believably, although I assume that these are going to be watermarked somehow by OpenAI, as the Nano Banana series of models by Google are. Ricardo Wolf here with a few more examples. Sam Alultman incoming car inside of the Tesla, you know, just to try some weird stuff. This dude is getting arrested. I mean, that looks very, very believable. Oh gosh. And the little subtle detail on the corner, you know, that's what I'm saying. Like you could tell, you could tell this is really going the direction that people are expecting and wanting from Image Gen as a whole. This very high ability, very realistic, believable at times. again. Hope hoping they don't labbotomize this thing fully. First person point of view going down the trail. Some dudes looking at you behind the counter. This is going to be a pretty big week for AI, I think. Thank you so much for watching, guys. Subscribe if this information was useful and helpful. Happy Horse I think has me a little bit more excited for some reason right now than the GPT image too, but I think it's just because of that opensource promise. And we'll see when it is released if they've set it up well already to work with more consumer grade hardware or not. Like we don't know a lot of the details. So like Happy Horse is more like mysterious. And the names got me. But this GPT image too, man, this this looks like it could be a true nano banana takedown model. Best image generation we've ever seen. I can't wait to try it on more complex stuff like comic book panels, some of my classic Nano Banana 2 prompts that I've been messing around with, YouTube thumbnails, stuff like that. GPT image gen has been decent, but it's never really caught up, I think, to what all the different providers and especially Nano Banana brought to the table. So, pretty great stuff. I'll see you guys in the next video and goodbye.