What’s new
- The foundation model behind Z-Image-Turbo: same 6B parameters and Single-Stream DiT architecture, but it keeps the full 50-step generation process for more diversity and controllability.
- Negative prompts actually work here: Turbo's CFG-free distillation weakens this, but the base model supports proper negative prompting for fine control.
- Built for downstream development: Tongyi Lab positions it as the base for community fine-tuning, LoRA training and ControlNet work.
- Same bilingual text rendering, different trade-off: visual quality is rated "High" (versus Turbo's "Very High") in Tongyi Lab's own comparison, trading some raw per-image quality for more diversity across runs.
What it’s good for
- Developers who want an open base to fine-tune, train LoRAs on, or use with ControlNet.
- Creative work that needs fine control over unwanted elements via negative prompts.
- Batch creative generation where you want more variation across outputs from the same prompt.
Prompting tips
- Use both positive and negative prompts: unlike Turbo, negative prompts are an effective control here — use them to cut flaws and unwanted elements.
- Keep guidance scale moderate: pushing it too high tends to distort the image; Tongyi Lab's reference range is moderate-to-low (3.0–5.0).
- Name the style you want: realism, anime, fine art — the model responds to a wide range of style terms.
- Enrich complex ideas before feeding them in: Tongyi Lab's own showcase demonstrates a "reasoning" style of prompting that lets the model draw on broader world knowledge rather than just parsing your words literally.
An old fisherman sits in a wooden boat mending a net, backlit silhouette at dusk,
sunlight glittering on the water, fishing boats heading home and a lighthouse in the distance,
realistic photography style, warm orange tones, grainy film texture.
Negative prompt: cartoonish, over-sharpened, malformed hands, blurry.
Known limits
- 50-step generation is much slower than Turbo, so it's not suited to real-time or low-latency interactive use.
- Editing is handled by a separate Z-Image-Edit variant — this base model is aimed mainly at text-to-image.
- Tongyi Lab's own model-zoo table rates the base model's visual quality as "High," a tier below Turbo's "Very High" — if you just want the best single image and don't need fine-tuning or diversity, Turbo may be the better fit.
Using it in Nomi
On the generation canvas, add a image card, pick “Z-Image” in the model picker and choose a mode (Text to image). Add your references, write the prompt and generate. In the storyboard you can pick it per shot. The first time, connect any one of the providers above under Settings → Models.
