Its been a while since I jumped into the AI Toolkit by Ostris and there are so many new models that have been released. So I decided to jump back into it to train my Desi Babes (for flux) on Z-Image Turbo using the toolkit.
My default setup for running this is to go into RunPod and subscribe to a RTX5090 which only costs $0.89 per hour which is rediculously cheap so I don’t bother stressing my RTX4080 locally and suffer with Out of Memory issues.
Ostris has released his entire app as a docker image which is available as a template on Runpod. Simply search under Templates and key in ostris, you will find it.
Once you run up your pod, launch the toolkit. Naviage to Datasets and create a new dataset, upload your images and caption them. If you need help with captioning check out some of my videos to help you with this. Search my ComfyUI Workflow repository for “caption” workflows.
Create a new job and set this up with following settings, many of them are default. If you are using VRAM more than 24GB then you can keep the identical settings.
Key changes I made: Turn off – Low VRAM, set Quantization Transformer to None, set Save Max Step saves to Keep to 8 (I like to have more files for testing), Turn on – Training > Cached Text Embedding. Rest of the settings are default.
Now you are ready to start the job and let it run. With the RTX5090, I was able to have the training done within 1 hour and 10 mins. The checkpoints were available and I could download the various versions, as a set it to save 8 different versions, I had those and the final version at 3000 steps.
Once you download these you can place them in the models\loras folder in ComfyUI and then you can start testing the different versions.
Here are the results of the training. Pretty happy with the results. I used the LoRA strength of 0.8-0.95.
If you'd like to support our site please consider buying us a Ko-fi, grab a product or subscribe. Need a faster GPU, get access to fastest GPUs for less than $1 per hour with RunPod.io