Coloring Book Z

Coloring Book Z is a Z-Image Turbo LoRA trained on a 100 image synthetic dataset that I personally generated, the images in the dataset were mostly human, vehicles and animal illustrations. All of these images were captioned using Joy Caption Batch.

Trained for 2,000 steps with Adafactor scheduler and a LR of 0.000075 using ai-toolkit with rank set to 6. For such simple styles less is more and I probably could have gone even lower to 4.

I'm attaching workflow examples to the images in the gallery, drag and drop into ComfyUI to use them. I recommend a strength of about 0.7, I used res_2m sampler, but it worked fine with euler as well.

Something I noticed is that the trigger for most of the Z-Image LoRAs are kind of redundant and sometimes confuse the model, just describing a black and white cartoon was generally enough to trigger the style. Adding words like simple or cute helped to simplify the look, experiment with asking for 3/4 angle and side profile views for some variety. I'd advise against using the word book as the model seems to interpret it literally.

Just a reminder that this is a Turbo model so expect quirks, i'll definitely share an update when the base model is released.

renderartist.com