Hey everyone, welcome back to my channel. So today is a day that everyone has been waiting for. Z image is finally released.
So we loved the Turbo Distilled model and now the full release of the model is here. This is a huge day in the AI space, especially for open source users. Uh we have a lot of hope for this checkpoint.
Um, and we're hoping to develop a lot of Loras and realistic content from it. So, this came out about 20 minutes ago. So, I'm recording this video just doing some preliminary testing.
Uh, so this is going to be very raw on the fly. No cherry-picking. Uh, I'll have a longer, better video later on once we know more about the model, what it responds to, Laura training, that kind of stuff.
So, you can get this on hugging face, but Comfy UI actually already has a template built in. So we'll show you that in about 2 seconds here. I just want to show you a couple of things on the Z image repository which is again hosted on HuggingFace.
So here's some examples they have. Looks very similar to what we know and love from the Zimage Turbo model. So the biggest thing is going to be the increase in diversity.
So if you use Zimage Turbo before, you probably understand that it was kind of hard to get consistent uh variations with prompting. you had to use Laura's or you had to be extremely creative with how you prompted things because often a prompt would give you the same image over and over again without any variations unless you added uh something to change it up. So this is the full model.
Hopefully it gets rid of that a little bit. Something I did see here that I found interesting was the visual quality was actually lower on the full model. I'm not really sure why.
Um, I'm going to assume they just cherrypicked some of the best parts for Zimage Turbo, but yeah, I did a couple of tests right before starting this video and it runs really quickly. So, they do recommend 28 to 50 steps. Uh, and you can also have a negative prompt now, which you couldn't have before.
So, let's not waste any more time. Let's move right over to Comfy UI and give it a couple of tests. So, this is one that I prompted right before I started this video.
So, again, trying to copy one of my drift girl Laura's. This is one of my prompts for my prompt library. So, I think this looks pretty good.
As I said before, you get to use a negative prompt now. So, that's kind of cool. Uh, what I did was I took the logic from that Z image announcement on hugging face and I plugged it into an LLM and I asked it to give me prompts based on how Zimage, the full model, likes to be prompted.
So, this is my best attempt at getting some good results. Now, it does recommend 28 to 50 steps. I did try a couple of generations, and I didn't see any markable quality improvements with the 50 steps.
I found that 30 was good enough. So, we're just going to stick with that for now. I may show you a comparison between the two.
So, let's get a new prompt. So, here is my list from chat GPT. So, I'm going to copy this one in.
And we will also take the negative prompt it generated. So, let's give that a try. You can see how fast this goes.
I won't speed it up. We'll just do this on the fly. Now, the model is already loaded, but this shows you what a live generation looks like once the model's loaded.
So, that's incredibly fast for a 30-step model of this quality. Okay, there you go. Now, I really like the way this looks.
The lighting is fantastic. You can even see a little bit of the small chin hairs there. This looks really, really, really good.
Uh, just to compare, I'm going to do the same prompt. We're going to try 50 steps this time. This one I will speed up just so we can save some time.
But again, it's incredibly fast. I mean, you can see it loading right now. Uh, this has got to be one of the more speedy models that exists.
Okay, so you can really see the diversity coming into play here. That was the exact same prompt with 50 steps. Uh, you can tell me which image you prefer.
I don't really see a ton of difference between the 30 steps and the 50. Uh, maybe you like this model a little more. I'm not sure.
One thing I'm noticing is this is incredible detail in the cloth on her scarf here. That's really cool. And again, I mean, other models can do stuff like this, uh, especially the closed source ones such as Nano Banana or even Chat GBT's image model, but this is open source and free.
Don't forget that. and it's running very fast on my 5090 and system with 128 GB of RAM. Now, everyone's here for AI influencers.
So, let's do an influencer prompt. And we're going to move this back to 30 steps cuz I don't see the need for 50. Okay, let's run that.
Okay, so this is our first attempt with three people in one photo. And I think it looks okay. I do see a few imperfections here.
Uh if you look at her lip, that looks a little strange. Same with that one. And the hands are a bit odd.
Uh let's run that one more time with 50 steps just to see if it makes a difference. It could be our prompt. Uh we're not fully educated on how to prompt yet.
So let's try that again with 50 steps and see if it gives any improvement. Okay. So I find this to look worse.
So it's definitely our prompt that's the problem. But as I said, this is just a live raw test. We're just seeing how this goes.
There will be much more detailed videos uh later in the week. I just wanted to get this one out just to show people what Z image is all about. But we're going to try one more here and see how it goes.
This is another drift girl type shot from my prompt library. And if you want to generate your own prompts uh based on photos, you can head over to smallzero. io and then you're going to go to the premium tools and then there is a tool called prompt alchemy that I host.
So this will help you generate prompts. You can upload a photo to get inspiration and then it'll spit out a prompt here. So that's smallzero.
io. You can also join our discord for free and our community. So head over there if you want to join the small zero community.
All right. So that might look familiar. We have generated something like this before.
Uh I think that looks pretty good. Again, we will need to tweak it a little bit. Uh the face is a little bit smooth.
I'm not loving the plastic detail on the car, but overall it looks really cool. So, one more thing I wanted to test was whether or not Laura's that I trained from Zimage Turbo would work right out of the box. Uh, as you can see on the screen here, I modified the workflow a little bit just to inject one of my Zimage Loras from before.
However, it was supposed to be a facial consistency Laura, but we got some very abstract looking art instead. So, looks like we are going to have to retrain our Loras. We'll do some more testing, but as of now, it looks like the Zimage Turbo Luras that you trained are no longer useful, which is unfortunate because we did think that they would transfer over, but I guess not.
All right, guys. That's all I wanted to show you today. There's going to be more videos coming, and you can join our Discord community for free.
If you want to get some tips and tricks, there's going to be people discussing this all day in there, I assume. Uh, just head over to smallzero. io.
If you want to support the channel, you can. It's just 10 bucks a month and you get access to the premium tools. Uh, otherwise, stay tuned.
We're going to have a lot of videos on this, I'm sure. All right, guys. Have a good day.
You can't see that. You can't see that. I'll be back.
Man, this gets deeper than a devil's leg in a dual sport. I stop the pedal to the metal break. Chains full of VS1's but the metal fake.
I'm f to drink a whole pint. That's how I meditate. Party on the hot block.
Hey everyone, welcome back to my channel. So today is a day that everyone has been waiting for. Z image is finally released.
So we loved the Turbo Distilled model and now the full release of the model is here. This is a huge day in the AI space, especially for open source users. Uh we have a lot of hope for this checkpoint.
Um, and we're hoping to develop a lot of Loras and realistic content from it. So, this came out about 20 minutes ago. So, I'm recording this video just doing some preliminary testing.
Uh, so this is going to be very raw on the fly. No cherry-picking. Uh, I'll have a longer, better video later on once we know more about the model, what it responds to, Laura training, that kind of stuff.
So, you can get this on hugging face, but Comfy UI actually already has a template built in. So we'll show you that in about 2 seconds here. I just want to show you a couple of things on the Z image repository which is again hosted on HuggingFace.
So here's some examples they have. Looks very similar to what we know and love from the Zimage Turbo model. So the biggest thing is going to be the increase in diversity.
So if you use Zimage Turbo before, you probably understand that it was kind of hard to get consistent uh variations with prompting. you had to use Laura's or you had to be extremely creative with how you prompted things because often a prompt would give you the same image over and over again without any variations unless you added uh something to change it up. So this is the full model.
Hopefully it gets rid of that a little bit. Something I did see here that I found interesting was the visual quality was actually lower on the full model. I'm not really sure why.
Um, I'm going to assume they just cherrypicked some of the best parts for Zimage Turbo, but yeah, I did a couple of tests right before starting this video and it runs really quickly. So, they do recommend 28 to 50 steps. Uh, and you can also have a negative prompt now, which you couldn't have before.
So, let's not waste any more time. Let's move right over to Comfy UI and give it a couple of tests. So, this is one that I prompted right before I started this video.
So, again, trying to copy one of my drift girl Laura's. This is one of my prompts for my prompt library. So, I think this looks pretty good.
As I said before, you get to use a negative prompt now. So, that's kind of cool. Uh, what I did was I took the logic from that Z image announcement on hugging face and I plugged it into an LLM and I asked it to give me prompts based on how Zimage, the full model, likes to be prompted.
So, this is my best attempt at getting some good results. Now, it does recommend 28 to 50 steps. I did try a couple of generations, and I didn't see any markable quality improvements with the 50 steps.
I found that 30 was good enough. So, we're just going to stick with that for now. I may show you a comparison between the two.
So, let's get a new prompt. So, here is my list from chat GPT. So, I'm going to copy this one in.
And we will also take the negative prompt it generated. So, let's give that a try. You can see how fast this goes.
I won't speed it up. We'll just do this on the fly. Now, the model is already loaded, but this shows you what a live generation looks like once the model's loaded.
So, that's incredibly fast for a 30-step model of this quality. Okay, there you go. Now, I really like the way this looks.
The lighting is fantastic. You can even see a little bit of the small chin hairs there. This looks really, really, really good.
Uh, just to compare, I'm going to do the same prompt. We're going to try 50 steps this time. This one I will speed up just so we can save some time.
But again, it's incredibly fast. I mean, you can see it loading right now. Uh, this has got to be one of the more speedy models that exists.
Okay, so you can really see the diversity coming into play here. That was the exact same prompt with 50 steps. Uh, you can tell me which image you prefer.
I don't really see a ton of difference between the 30 steps and the 50. Uh, maybe you like this model a little more. I'm not sure.
One thing I'm noticing is this is incredible detail in the cloth on her scarf here. That's really cool. And again, I mean, other models can do stuff like this, uh, especially the closed source ones such as Nano Banana or even Chat GBT's image model, but this is open source and free.
Don't forget that. and it's running very fast on my 5090 and system with 128 GB of RAM. Now, everyone's here for AI influencers.
So, let's do an influencer prompt. And we're going to move this back to 30 steps cuz I don't see the need for 50. Okay, let's run that.
Okay, so this is our first attempt with three people in one photo. And I think it looks okay. I do see a few imperfections here.
Uh if you look at her lip, that looks a little strange. Same with that one. And the hands are a bit odd.
Uh let's run that one more time with 50 steps just to see if it makes a difference. It could be our prompt. Uh we're not fully educated on how to prompt yet.
So let's try that again with 50 steps and see if it gives any improvement. Okay. So I find this to look worse.
So it's definitely our prompt that's the problem. But as I said, this is just a live raw test. We're just seeing how this goes.
There will be much more detailed videos uh later in the week. I just wanted to get this one out just to show people what Z image is all about. But we're going to try one more here and see how it goes.
This is another drift girl type shot from my prompt library. And if you want to generate your own prompts uh based on photos, you can head over to smallzero. io and then you're going to go to the premium tools and then there is a tool called prompt alchemy that I host.
So this will help you generate prompts. You can upload a photo to get inspiration and then it'll spit out a prompt here. So that's smallzero.
io. You can also join our discord for free and our community. So head over there if you want to join the small zero community.
All right. So that might look familiar. We have generated something like this before.
Uh I think that looks pretty good. Again, we will need to tweak it a little bit. Uh the face is a little bit smooth.
I'm not loving the plastic detail on the car, but overall it looks really cool. So, one more thing I wanted to test was whether or not Laura's that I trained from Zimage Turbo would work right out of the box. Uh, as you can see on the screen here, I modified the workflow a little bit just to inject one of my Zimage Loras from before.
However, it was supposed to be a facial consistency Laura, but we got some very abstract looking art instead. So, looks like we are going to have to retrain our Loras. We'll do some more testing, but as of now, it looks like the Zimage Turbo Luras that you trained are no longer useful, which is unfortunate because we did think that they would transfer over, but I guess not.
All right, guys. That's all I wanted to show you today. There's going to be more videos coming, and you can join our Discord community for free.
If you want to get some tips and tricks, there's going to be people discussing this all day in there, I assume. Uh, just head over to smallzero. io.
If you want to support the channel, you can. It's just 10 bucks a month and you get access to the premium tools. Uh, otherwise, stay tuned.
We're going to have a lot of videos on this, I'm sure. All right, guys. Have a good day.
You can't see that. You can't see that. I'll be back.
Man, this gets deeper than a devil's leg in a dual sport. I stop the pedal to the metal break. Chains full of VS1's but the metal fake.
I'm f to drink a whole pint. That's how I meditate. Party on the hot block.
