HiDream-O1 Explained: Why This AI Image Model Feels Different

Ajouté : 2026-05-11

[00:00:00]Holy buckets of bird fat. It is the new image generator that is out. Hydream 01 is here now. Uh 01 Hydream has been around for a little while. They came out last year. Um and it was actually going to rival Flux there for a while, but Flux just too difficult to um knock off at the time. That was until of course Zage came out. Now with this new multi-modal model which means that it is a one model does it all sort of thing.

[00:00:33]You can edit images you can create new images do all sorts of thing do it fast with the higher level things guys if you can get this running on a bigger machine this could be rivaling eventually what's coming out with chat gpt and its images as well. So let's take a look at it right now. Hell.

[00:01:02]All right. Well, welcome to Get Going Fast. I am Gary, the AI hobby guy. And welcome. Uh, you may realize I rebranded recently as the AI hobby guy. And that essentially is just a way that's going to open up the door to be able to do more sort of stuff. Actually start traveling to visit you guys. Um help on your sort of projects. Really expand what we have here. Contining continuing to do what we do and what we do well.

[00:01:28]Teaching you guys how to install, giving you a place that um you could feel confident about the information you're getting um and show you how to do it.

[00:01:36]Right. Hey, but we're also going to expand this because I want to actually start going out and uh meeting with you guys in person and working on your systems as well. So anyway, welcome to Get Going Fast. I am the hobby guy. You found the good place. All right, so um we're looking at Hydream here. Now, Hydream is a model that came out a year ago. In fact, let me show you what it looked like back then. Okay, if we head over here, this is that get going fast.

[00:02:04]um grow the members site. Um you can see these are the kind of images we were making before. So at the time these were really awesome. Okay. Um obviously these aren't as awesome now, but does show you how far [laughter] the uh um the hobby has come. But the point is this was the original and like I said it came out um it was hey the flux killer the flux but flex was too strong. It was like these two models got into this the actually it was like these two models got into a a ring full of jello- and they wrestled around.

[00:02:35]Everyone was like yeah like like right um at any rate um flux one we didn't hear anything about Hydream until now. Okay. But now Hydream's come back all the all the sexier looking ready to uh for round two and it's looking good.

[00:02:53]So, as I was saying, we're take a look at some of these images. Um, Ocean, this is an FP8 model here that I'm using. So, you got to think um you have a full model and then you half it and then you half it and you half it like again kind of a thing. And that's what this is.

[00:03:10]This is like a half of a half of a half um kind of thing. And that's still pretty good. The more you look at it, you can see that there's some details that aren't uh showing up, but that's pretty good. And if you can get this thing to run at higher levels, it's going to be even better. I mean, again, here, look at this. Generated this. This is um a great thing. Now, obviously didn't get the Ninja Turtle all that fantastic, but this is a great um image, but there's something more about this model that is what actually really makes it so powerful because like I said, this this is pretty powerful. And I will tell you using Comfy UI with this. Yo, I was dropping these images in about 20 seconds and that's48 by 20048. That is fullon high definition. Okay, so these images that I showed you over here, okay, these ones here, all this stuff, that's like 20 seconds a thing. That but that's the FP8. Okay. Now, I have another app that that we've put together another one here. You're or with the same workflow, you can use the FP16 or the B, you know, the BF16. So, that's like twice the amount of this, right? 16 divided by two is eight. So, eight is half the size of the 16. Um, but that one will put out even sharper stuff for you, I'm sure.

[00:04:30]But, but I know what you're asking.

[00:04:32]You're saying Jerry AI hobby guy, you told us you was going to tell us what's so big about this. Okay, I'm going to do it. Just hang here. Watch this. And >> so to understand what's different about this high dream model, what we really need to do is we kind of need to understand the concept of how AI models are designed in the first place. So typically we have a neural network.

[00:05:03]Okay. So with the software we create a brain and that brain is trained on particular things whether it's speaking whether it's creating an image or doing whatever it has one thing that it really knows how to do. It'd be like we take it we put all the memories in it. We put all the neurons we say this is you already know this stuff and then we birth it. WE BRING WE'RE LIKE IT'S ALIVE AND IT COMES alive and it already knows the things it knows. It doesn't have to learn them because it it's just there.

[00:05:31]It'd be like if you suddenly just appeared and you have these memories but you didn't actually live them. They were embedded in you. So that's essentially what we're doing with AI. Now how these models, these image models typically work is you have several AI brains that are working together. This is where things like the concept of our text encoder comes in or our VAE. We have different models that are trained to do different things. And so we start out say with our image model and we do as much as we can in there because it's trained, its brain is trained on that one thing and it says, "Okay, now I'm going to kick it out over here to this other brain and I'm going to have this brain do some stuff." Okay? And it's going to go over here and and then it's going to go to the text encoder and the text encoder will do some stuff and then we're going to we have to keep these kind of three things synced together.

[00:06:21]Yo, and then we'll pull it all back together into one glorious image because AI is is magic, right? Okay. So, that's great. And that's how most models work.

[00:06:33]Okay. But it's also why they take time because uh it'd be like throwing three people in the room. You got three. You want to put something in, you get three experts. And so, you got to get the one expert to agree with the second expert who to agree with the third expert. And you got to get all of them kind of in line on the same page. working together.

[00:06:52]Okay. So, what does Hydream do? Great question. Yo. Okay. That was that's my chat GPT impersonation by the way.

[00:07:01][laughter] What what does it do? Great question. That was that really insightful. Okay. So, the what's the difference? The difference is when we pop over to the single model, this hydream, it has all of that baked already into the brain. So, you don't need three models because essentially, watch this. I'm going to make this bigger. Boom. Boom. Essentially, we just made a bigger brain.

[00:07:27]I mean, we didn't do it. You know, the smart people did. I mean, you were smarter. You know, the other smart people, not me for sure, but it makes a bigger brain, right? Say like the the neural the amount of connections it has is like this size, right? It's like this size. And then we say, well, actually, let's go ahead and make it so it has everything. And we'll make the brain like this. So, not only are we going to add uh memories about the childhood and this and that, but we're going to add memories about the, you know, say we were creating something with fake memories and we say we'll also create memories of a job it never had and we'll create memories of skills that it never had. You know, this kind of thing and then it activates and goes, "Oh, I know all these things."

[00:08:06]So, Hydream knows how to do the work of the VAE. It knows how to do the work of the text encoder. It knows how to do all these things. So, it can do all of it there without having to pass it around.

[00:08:17]Imagine if you knew enough that you don't have to have those other two schlobs in here the room with you, right? You're sitting there like, "Oh my god, we we can't even agree on lunch, let alone getting on the same page. Get rid of them. You know what? I've got enough of the information in my head. I can do this myself." That is Hydream.

[00:08:35]That's why hydream is such a big deal or one of the reasons why hydream is such a uh a big deal. One other great thing is happening with this model and that's it use what's called pixel space. Now to understand pixel space you have to understand what most image models do. What they do is they create an image, right? And then they take it down to what's called a latent space and they make changes and then they bring it back up to the image and they go, "Okay, I need to make this change." And they bring it back down to what's called the latent space and they make changes and they go back up again and they kind of keep doing this. Now, there's a conversion that's taking place because you're compressing these this this image, right? compressing it, making changes on on a compressed level, and then you're expanding it back out again. So, you can lose data. Now, what Hydream does is Hydream says, "We're not going to bother with the compression stuff. We're going to work with the image itself." So, that way, we're not worried so much about losing consistency. Right? See? So, so I I've got this magic wand and I and I've got a blueprint of a house and I wave it at this blueprint and I go bam and that blueprint goes poof and all of a sudden there's a big house, big white mansion house. I'm like, "Oh man, I love it.

[00:09:54]This is so cool." You know, the tennis courts, all this, but oh, oh man, there's no swimming pool, right? So, I wave my wand at the house and it becomes a blueprint again. And then I work on the blueprint and I change. See, it went from like a a big object down to an abstracted object. the blueprint. I make changes on it and then I go, "Whoa, wait, my wand again." And then poof, the house is there. There's the swimming pool. All this kind of stuff. Okay. So, there's a process that takes place and and maybe when it comes back, there's some little small changes, too. All right. What Hydream is like says does is it takes the house and instead of wearing with a blueprint, it says, "Let's just make changes to the house itself." So, so I don't even need to abstract it. I'm just going to work with it as it is. Right? So then I can just walk around the corner and I'd be like, "Okay, I'm going to put the pool right here, right? And we'll put the pool in there." I don't got to worry about compressing the basketball courts over here or any of that kind of stuff. I just going to here. So in the same sense, right? The first one you're compressing, bring it down to something and then bring it back. Hydream doesn't do that. it's working closer to the image, which means less mistakes, more consistency, more coherent um imagery at the end. And lastly, what would end up making this model fantastic uh is the simplicity of training it can again. So people that are wanting to fine-train these things, right? They um typically you have to get all you got to get all your duction or you got to get all you got to train it on the right VE the train text all this kind of stuff.

[00:11:37]Well, because it's all baked in all you have to do is train the one model and um without having to worry about putting all the pieces together.

[00:11:47]Again, the idea being like if you have a bunch of things working together, you got to make sure all the little pieces work together. when you have one model, there is no little pieces because it's all just one piece. Okay, so that's how Hydream works. Hopefully that was a uh um good explanation for you. You can get the model manager over here at getgo going fast.pro and what this will do is actually you'll just install it into your Comfy UI. It will actually um it has to download some custom nodes. Um it will do everything it needs to do to your Comfy UI. Okay. Then it'll load it up and will allow you to download um the model by yourself, which is uh which is really cool. In fact, I'll show you here really quickly right over here. Uh you'll see here you just pick one of these and then click the run. I actually have to check this download if missing and it will actually download it for you. Now, you need to use the the dev FP8. Okay. I don't think you're going to be able to run the others if you have a bigger machine. I do have a standalone app that does not come for UI and it will run the FP16 as well. Um, again, I got a friend who's got a smaller video card than I did. He says with the F the BF16, he's popping these out in a minute. I got a 4090 and I mine's not going that fast. So, I don't know. Um, maybe it's probably something with my machine because I do a lot of stuff. So, maybe mine just need to be cleaned out. But the point is, you go to getgoingfast.pro, you can get the comfy UI installer. That sucker's fast.

[00:13:18]20 seconds, yo. To get a thing, or um you can get the standalone, which is up there as well, and you can play around with that. All free for members. All right. Well, we're going to wrap it up right there. Thanks for uh hanging out with us. If you did like the video and it was helpful, go ahead, hit that subscribe button and tell other people about it as well. You can always send your thanks through a little thank you note down there. Uh send a couple bucks in my pocket as this is my business. So you're helping the hobby community move forward. Um as usual, you know, have a great day. It's still Mother's Day. Call your mom if you forgot to call her. Just call her tomorrow. Let her know you love her. And um you know, oh, come hang out with us on the Discord as well. We'll be live tomorrow for live stream. So come check out with us with that. Oh, that's always fun. weekly review. And with all that said, I'm going to give you one of these. I wish you the very best.