To make this clear, I ran the same set of 60 images from my dataset on both models. Flux dev has great similarity and follows the prompt steadily, but the files are big, and on my 4090 it takes 12 seconds per image.
Z-Image Turbo, on the other hand, matches the similarity, sometimes even has better skin texture, and training time is cut in half—4 seconds per image. The only catch is it’s really unforgiving with bad training data; Flux can handle dirty data, but this one will throw a fit.
I generate in bulk, so the speed advantage adds up fast—I just moved my whole workflow over. Anyone else made the switch? Let’s chat about your experience.