I fed the same 60-image dataset into two models to train a character. Flux Dev started looking pretty similar around epoch 12, but the files are huge and each image takes 12 seconds. Z-Image Turbo matches it on likeness, sometimes even has better skin texture, training time is cut in half, and generating images takes 4 seconds.
But it has a clear weakness—it’s really unforgiving with dirty data. If there’s a bad sample mixed into the training set, it messes up, while Flux handles it better.
For people who generate in bulk, speed adds up over time—saving a few seconds per image becomes huge when you’re doing thousands, so I just moved the whole pipeline over. Wondering how many others working on characters have switched to Z-Image.
