Z-Image Turbo Lora Training in Ostris AI Toolkit

Z-Image Turbo is an amazing new image model that was just released by the team at Alibaba. It runs

well on low-power GPUs and produces phenomenal results. While the open-source community is eagerly awaiting the base model to come out, which will likely produce even better results when training Loras and checkpoints, here's how you can train a Lora with this distilled model today.

To get started, I'm going over to RunPod, and I'm going to deploy an RTX 5090, which is sufficient for training this model. Now you want to make sure that you change out the template to the Ostrais template by simply searching for it within the search bar. Make sure it's this Ostrais AI Toolkit latest version. Click that. Ensure you have

bar. Make sure it's this Ostrais AI Toolkit latest version. Click that. Ensure you have enough space on the disk; the default should be fine. Then just simply click Deploy on Demand.

All right, so here's the Ostrais AI Toolkit. I already have a job running, but we'll go ahead and create another one. Simply click on Datasets and then click on New Dataset.

We're going to give this dataset a name. I'm just going to give it a generic name here, so I named it Teacher with a 3 as the second-to-last character, and then let me bring in some images.

These nine images I used to train a FluxLora, so we're just kind of reusing these ones. You

can add a caption to the dataset if you want; however, I'm not going to worry about that.

Now we're going to create a new job. Let me just set the training name and the trigger word. As

for the model architecture, we'll want to select Z-Image Turbo with the training adapter. You'll

see the training adapter path. I just want to quickly note that the creator of this application mentioned they are testing out a new training adapter. If you want to test that, say within the next week or two, just simply change this out from V1 to V2.

As far as the datasets, we'll change the target dataset to this teacher one. Regarding the prompts that we can see of the generated images while it's training, you can set those as you want. Let me go ahead and do that. I'm basically creating two prompts that we can test out here. All right,

those prompts are set. All I need to do now is simply click on Create Job. If

I go over to the training queue here, we'll see the job here, and then we can press Play and then click Start. It will go through the process of training all the images.

I found that with an RTX 1590, it roughly takes about an hour or so to generate a Lora at 3,000 steps with all the default settings. Once a job is complete, as we can see here, if you click on Samples, you can see how the model started to learn the concept over time.

So here's the original prompts with the Lora not really being applied, and as you can see, the characters slowly morph into my teacher illustrated here and look pretty good.

Now we can actually download this file. If I go to Overview, then to the checkpoints here, you can download this latest checkpoint and use it within your workflows. As you can see, I have the Lora loaded here, and I have a slightly different trigger word,

but it's the same exact dataset and settings as I went over just previously. You can see it's a schoolteacher shooting a basketball and smiling, and the output looks phenomenal.

That's it for training a Z-Image Turbo Lora. If you have any questions, please drop them in the comments section below. Thanks so much for watching, guys, and I'll see you in the next one. Take care.