MiniMax H3 Is Taking Over AI Video and It's Open Weights transcript

Open source LLM surprising us a lot this month. They're taking over general purpose, comparing and sometimes beating top models such as Fable 5. And now it also comes to video. We have an amazing Minimux H3 released with open weights already. Community absolutely love it. It's already integrated natively in Comfy UI. And it has some unique features that paid model doesn't. We're gonna run amazing tests and I'm also gonna show you how to run that locally.

And share some tips. So hit that like button to support what I'm doing. What do we know about it? It supports variety of inputs such as text, image, video and audio. And it's something unique about audio because it's not only taking audio as reference, but can build video on top of audio, which is quite a unique thing. Resolution up to 2K and duration up to 15 seconds. Also about how audio is generated and that it goes in one single pass.

That should mean that audio must follow better the video content. Well, that's very interesting. We're gonna do some tests related to that. Absolutely. So let's dive into it. Most of my tests I did in Minimux design, which is an official app and probably the cheapest way possible to access newest Minimux H3, which will cost you only six cents per second of 2K resolution. Well, that's so much cheaper than C Dense or any other models.

And that's also cheaper than using a true Halu website. So let's start with the first test. We're going to also do some complex tests for different topics, but I want to open the floor with some anime tests because I saw something like this on X and I was like, no way you can do exactly like that. And this one is like anime PV with nice, cool transitions, like imagine Gacha game where you have new character preview. So here's the reference image.

Here's the prompt. So the workflow again is pretty easy, especially in Minimux design. You have the agent that will help you out. But you provide the prompt, you set the tags, whether that's character, shit, etc. All the prompting I do with the AI guys. No real magic. You just need a good idea. Then agent going to cook for a while. I also use that skill called anime game PV. And by the way, cool thing is that in Minimux design skills are triggered automatically when you need them.

You just mentioned that you need something like that and it will be there and agent starts to cook. So sometimes it can ask some questions, clarify some stuff. You can see all that, change race music, voiceover, etc. And here's the result. And dang, I was blown away, really, because this is not what I expected. It will be doing good, right? I mean, maybe more realistic frames, etc. But this is some cool stuff that you can drop in your game, especially if you do anime games.

I'm super hyped about this particular skills. And later in this video, you can see that Minimux surprisingly good in some similar stuff. Like product showcases, some cool slow motion stuff, etc. I did one more test with anime and it looks also very dope. You can see the prompts and I promise I'm not sure, but maybe we will do more anime later. But this one is also very cool, especially this transitions. I like how sharp they are.

And since I mentioned product showcase, here's one more prompt. And I can actually see how this is used commercially. It's absolutely cool. Dang. Prompt is on your screen. You can see what I requested, like, and what I got. I'm not saying like it's flawless, guys. Of course not. And for some prompts, it's not one shot. It's like sometimes like to get best results, you're going to run it like one to time. And now more about audio, because that's sort of like a unique feature of Minimux.

What do we have? So let's do some tests. Like, let's say I have this audio. It's a simple melody that I wanted to make a video that's going to follow that. Sort of like equalizer, but nicer. So here you can see the prompts and some results that I got. Here's the first prompt that I tried. That was pretty much just text input. OK. And what I can say is like that the beginning was quite nice. But let's say second half of the video, it was not fitting that well.

I did one more with more details. And this one again, same thing like beginning. They're impressive. Like it really fits the beat. But somewhere out there, the middle, it starts to fit less. But my last try was actually very successful. Just look at this video. And I think that that's just super nice. Maybe because I gave it like more freedom in terms of like abstract figures, etc. But the last one, my absolute favorite.

Of course, we need to test it. How good it will follow the voice reference to evaluate that properly. Your humble servant created this humble character for educational purpose only and for bait prompt is on your screen. The prompt is a bit long, right? But it handles nicely. So let's take a look on the results. You said you'd call when you landed. That was Tuesday. I'm not angry. I'm just telling you the tide comes in at six and I'm not waiting.

Well, the lip sync is flawless, but the audio is very isolated. I wish it blends more into what's happening in the video. Right. So I can barely hear like the weight sound. I guess like not the best AI generated voice. So that one can definitely be improved. Another significant feature is so -called R2V. I don't know. Like if you scroll X, you can see a lot like R2V, Minimax R2V. But that basically stands for reference to video and means that you can drop like a reference of the camera, etc.

That's not like a new invention. What's the way Minimax H3 handles that? That's quite upgraded. So here I found this cool drone camera recording. And I'm going to use it to create some different, like say, ancient room or anything like that. And let's see how good I'm going to follow the motion. And not going to lie, it's quite good. It's quite good. Right. I did a few tests. You can see the second one. Pretty same. So that one actually works good.

I also tried some cinematic shots like this prompt where I wanted like two angles of one same scene so I can later combine them. But I would say for difficult cinematic action scenes, it's maybe not the best use case, at least right now. There are a few things that you need to consider. The max length is 15 seconds. So you're still going to need to think in segments. And even though it gives quite great results in 2K, but I spot that on very action paced clips, we can see some blurriness.

Like it's light, maybe not that significant, but it still exists. And I have to be honest about it. I think it was most visible when I tried my benchmark test with 3D. First of all, it was following 3D reference exceptionally, right? So here's my play blast. And then you can see the final video. Right. I mean, the process is pretty much the same. You provide the key references, etc. And you may see when this camera goes too fast, like there are some blurriness, but the slower it moves, like the sharper the image is.

So I just recommend to consider that to use Mini Max more efficient. But the opposite of it, that slow motion or some cool transitions are handled insanely well, and let me show you what I'm talking about. I don't know. I was just watching like some of Mr. Beast video and I decided like, can it actually make this motion graphics lag for some of the challenge? And there is not much sense, just some stuff happening around.

You can see the prompts. So let me show you the result of first prompt. And I was frankly surprised. And for the reference, I was not providing all these images. And here's another try with another prompt. Also look good. Like the beginning is a little bit strange, like when it's hide it like one part, but the rest again, still, it's very cool. And you can control that. That's amazing. Then I thought to test it, like how good it can handle consistency of text.

And here's the product and you can see very prominent text. Well, at least just so we can follow easily and if it breaks something. And you can see the prompts. Claude helped me to put all text from labels inside the prompt. So you can see it's also here. Let's see if it actually help. Here's the first test I did. And indeed, text is there. Can't say that, like I really like the ads, but it has something in it, especially the ending, right?

Boom. Gentle all day. Not all tries was really nice, but here's the last one that I really liked. Gentle all day. Gentle all day. Yeah, I wish I'd be gentle all day. Well, text are checked and that's really cool. I mean, that's expected, right? Like we do not expect gibberish because like all frontier, like not a banana or like gibberish image, they already can handle text nicely. In 3D generation, however, like it's still an issue.

I don't know what they use inside, but when we generate like 3D and do the texturing, like often it breaks text and generate gibberish. But here it's quite fine. And one ultimate test will be like, let's say we got brief from the client, like for a new product. And here I actually prepared that brief. Here you can see the PDF file, like we can see what's in there. Basically has quite a lot of information and there are deliverables outlined.

There will be a master video, cut down, a vertical video, etc. And that's a good challenge for this Mini Max agent, right? So what I'm going to do, I'm going to drop it all to Mini Max and attach the brief and images and just say like, make the film. Deliver everything in Section 9, like most lazy approach. So let this guy cook and let's see what it actually did. Here's the first 15 second video and it's freaking insane.

Wow. This is like watching normal commercial. OK, so something wrong happened here with this battery. Oh, wow. Like, OK, haha. Is that supposed to be like that? OK, the rest looks fine, I guess. OK, I like it. I mean, overall, it's pretty cool. I also generated short video. OK, wait, what's there? OK, some gibberish. Look, kila Mers 10 kila Mers, gibberish spots. Other than that, the text that goes on top. No questions, right?

No gibberish. Fine. And here's the video that it assembled like the full length. 30 seconds. Let's see. OK, here is some gibberish. Sorry for that. I like how it mixes like this motion graphics with the video. Oh, now it finally looks like normally, but I don't think it should be like that. OK, wow. Let me know what you think, guys. Of course, we are not publishing that right away, but some fragments of it definitely can be part of commercial.

They look just really cool. And I guess like if you do smaller iterations, like generated like in smaller segments, provide maybe more details and reference, that's going to be all good. I also quickly tested like languages and I have like mixed reviews here. So here's the prompts and the first one is in English. No sugar. No hurry. You can hear that girl announcing like the motor of the brand, I guess. Then we have the next one was supposed to be Spanish.

See, no sugar. Simpli, sir. And the text on the bottle was in Spanish. And then it's not in Spanish anymore. Well, she's speaking Spanish. Strange and Japanese again, should be Japanese. Muto, you could eat. So quite random, like in Spanish, it changes what's on the can. Then in Japanese, it didn't change what's on the can. Well, the voiceover was changed in both examples. And in the end, in Japanese version, there is like Japanese.

I don't know. Like if you're from Japan, let me know if that's true. Per sure not. Another test that I also saw a lot on X and definitely wanted to try is gaming TV with dialogues. And first I tried to do it with the play blast provided. And here you can see what I did, actually. I provided like these characters, this play blast and this big, big prompt. Explaining all that stuff with all the references. And to be honest, I was really not happy with the results.

I guess, like it was looking very, very AI, especially the movements, etc. and like the consistency. And what's happening was not really consistent. I tried a few times with slightly different prompts. It reminds me this gaming Instagram ads that I'm constantly getting watched. Really hate them. But hear me out. I decided to give it a try without play blast. Here's the prompts and here is the results. And this one is completely different thing.

Dang, I like this motion. Again, like few moments there are not really making sense. And I guess maybe it's not really, really exactly following the prompts. But like 80 percent of that can be used. And that's a really, really impressive progress. This one looks much better to me than the first one. One last test before we move to local setup, I tried like how good it will handle if I provided with all possible type of reference, like play blast, character reference, object reference, like voice reference and the prompt, etc.

Like the only thing I didn't provide it is like another brief, but brief basically still is text basically. Right. So you can see this on the screen. And here's the results. People ask me what makes it worth the money. It isn't the leather. It isn't the soul. It's that after a week, you stop thinking about your feet. And here is the first version with play blast lip sync. Perfect references. Perfect play blast. I'm not sure now.

I'm really not sure now if like it works really good with play blast, because without play blast, I got better results. People ask me what makes it worth the money. It isn't the leather. It isn't the soul. It's that after a week, you stop thinking about your feet like this one. And, you know, it actually understands better, like all these camera transitions than I used to when I was working like with seed and for example.

So unexpected. I guess I had to revisit my recommendation on using play blast. And by the way, if you want to try this play blast approach, actually in minimax, they do have 3D director stage, which is sort of like a blender replacement that you can access right here. So we can open that and you can see like there is a camera. There is objects you can add more objects, generate of them. Really convenient. Take a look if you need it.

Anyway, now how do you run this amazing model locally? Well, first of all, again, as I said, it's already released. Open weights are on hugging face. You can figure it out. And there are different versions like there's someone already made like turbo max version, like if you follow that thing, you can see that. But the easiest way probably will be through Comfy UI, which I actually did. And the good news, it's officially included in Comfy UI.

What does it mean if you go and do Comfy desktop and basically install it after installation, you can even see this choose a starter workflow. And here minimax is top one suggestion. And that's definitely the best local model you can run. There's also Seedance 2 .5, but don't be delusional. It's API. It's not what you can run locally. So we're going to choose minimax, hit install, and we're going to wait. We're going to wait for a while because it's 53 gigabytes.

And it will definitely take some time. And if you have 24 gigs plus, most likely it's going to work. Like you can check out on the settings. But for my setup, I was testing it on 1590 24 gigs. I figure out like the best setup that also speed up the generation will be the following. So in startup arguments, we're going to need to set this usage attention. Fast disk and cache none. That's some optimization that I figured out, like for cache optimization, like memory optimization and Sage attention, like speed up generation nicely.

And then we're going to need to go to terminal here and run this two commands. Pipe install Triton. You can see it on the screen and also install Sage attention. That's it. Then we're going to restart it and we will be able to access minimax workflows. You go to templates and you can see all this minimax are here. Image to video, reference to video, text to video. And if you need more confioy templates, you can actually find them in minimax design as well.

They just added like confioy workflows here. And you can see like there's plenty of them for different purposes and using confioy, just download them from here. That will basically show you the presets of this node setup with plenty of explanations, et cetera. So let's give it a test right away. Of course, I'm going to test it with Jojo. If you don't know, I'm a big fan. Of course, locally on 24 gigs, even though it's possible, it's going to be longer.

For five seconds, full HD video. It took me 15 minutes to generate. Well, that's the answer why, of course, I was doing all the tests on the platform. And to be honest, for the current price, that's crazy. Like that's really, really cheap. That was my croissant. But this is so amazing that these capable models are now open source. And again, like you can access them from consumers grade. Again, that's very important because like there are some big LLMs even though they're open weights, they're not fitting consumers grade GPU, like most of them, right.

But this one is like a top tier that fits consumers grade. I'm super hyped. And also for developers, I recommend to take a look on official platform that Minimax .io where they explores even more capabilities, such as video regeneration, a .k .a. upscaler and HR context or air API. And much more if you know what to do with that. If you're building any apps around video, I think that's frankly like the cheapest option right now from what I saw.

And guys, I'm testing these things quite a lot. And one last cool thing is that Minimax design that I was using probably also going to have local version that you can run locally using your own VRAM, which is absolutely cool. There is no such thing like except Comfy UI. But if that going to be the same way with the agent, like super cool to see that that was the as usual. Like and subscribe and I see in the next video.