Z-Image-Re-Turbo
This is a re-turbo (de-restoration + re-acceleration) edition built on top of the excellent Z-Image-De-Turbo weights generously shared by the original author – huge thanks for making De-Turbo public!
After extensive LoRA training on ai-toolkit, I noticed that the de-distillation process, while perfect for training, slightly slows down inference compared to the original Turbo model. To bring back the lightning-fast generation speed while fully preserving the training-friendly properties of De-Turbo, I performed a careful re-acceleration fine-tune on high-quality outputs from De-Turbo itself. The result is Z-Image-Re-Turbo: it restores near-original Turbo inference speed, yet still behaves exactly like De-Turbo during LoRA/training, maintaining perfect compatibility with the entire Z-Image ecosystem.
Usage (Inference)
Recommended: CFG 1, 8-10 steps, most samplers work great
CFG normalization and simple Euler/DEIS still give excellent results
Speed is now on par with (sometimes even slightly faster than) the original Turbo
Why Re-Turbo instead of just going back to the original Turbo?
Because I (and many others) have already trained dozens of high-quality LoRAs with Ai-toolkit. Re-Turbo gives you the best of both worlds: the freedom and quality of De-Turbo for training, plus Turbo-level speed for everyday generation. Once again, massive respect and thanks to the original De-Turbo author – this model literally wouldn’t exist without your groundbreaking work!