Z-Image

Z-Image is the foundation model of Alibaba's Z-Image family — the full-capacity transformer from which the eight-step Z-Image-Turbo was distilled. It was published second: Turbo's weights went up on 25 November 2025, this one on 23 January 2026, both on Apache 2.0 with no gating. Architecturally the two are the same 6-billion-parameter single-stream diffusion transformer; what differs is the post-training. Turbo was pushed through distillation and reinforcement learning to finish in eight passes, while Z-Image stops after supervised fine-tuning and keeps the complete training signal. That difference is the whole point of the release. Because it is undistilled, Z-Image supports full classifier-free guidance, which is what makes negative prompts work — a user can reliably suppress an artefact or a composition instead of re-rolling the seed. Alibaba also reports markedly higher variation between seeds, which matters in multi-person scenes where a distilled model tends to repeat the same faces, and lists the model as the intended base for LoRA training, ControlNet-style structural conditioning and semantic conditioning. Recommended settings are published rather than left to guesswork: 512×512 up to 2048×2048 by total pixel area at any aspect ratio, guidance 3.0 to 5.0, 28 to 50 steps. In Alibaba's own comparison table the trade is stated plainly: Turbo wins on visual quality, Z-Image on diversity and fine-tunability. A third checkpoint, Z-Image-Omni-Base, is listed as the rawest starting point of all, and Z-Image-Edit as an editing variant; as of 17 September 2026 both are still marked "to be released" in the model zoo, so the family that is actually downloadable is these two. Traffic follows the fast model: 63,475 downloads in the thirty days to 17 September 2026 against 681,390 for Turbo — a tenth of the traffic for the model that every fine-tune of this family has to start from.

▸Videos

No videos yet.

▸Reviews