What’s new
- 6B parameters, 8 steps to an image: distilled for speed — sub-second inference on an enterprise H800 GPU, and it still fits on a consumer 16GB card.
- Accurate bilingual text rendering: text rendering is a classic weak spot for open image models, and Tongyi Lab's own showcase highlights a clear improvement here in both Chinese and English.
- A strong benchmark showing: per Z-Image's own repository update on 2025-12-08, it ranks 8th overall on the Artificial Analysis text-to-image leaderboard and 1st among open-source models.
- Single-Stream DiT architecture: text, visual semantic tokens and image VAE tokens are concatenated into one sequence, which Tongyi Lab says is more parameter-efficient than dual-stream designs.
What it’s good for
- Interactive products that want near-top-tier quality with fast generation and low deployment cost.
- Local deployment on consumer GPUs (16GB VRAM or less).
- Posters and social media graphics that mix Chinese and English text.
Prompting tips
- Be specific, and don't worry about length: Tongyi Lab's own showcase examples get more stable results from more specific detail, not less.
- Write bilingual text straight into the prompt: whatever Chinese or English text you want rendered, type it as-is — this is one of the model's strong points.
- Name the style explicitly: ask for photography-style description for realism, or specific style terms for anime/illustration — the official showcase spans everything from photorealism to anime.
- Switch to the base model for fine negative control: Turbo uses CFG-free distillation, so negative prompts have limited effect; for precise exclusion of specific elements, the base Z-Image model does better.
A cyberpunk city-night poster, neon shop signs reading "Future Market" in glowing letters,
rain-slicked streets reflecting colorful lights, a silhouette holding an umbrella at the center,
high-contrast purple-blue and pink color palette, detailed illustration style.
Known limits
- Speed comes at the cost of diversity and fine-tunability: Tongyi Lab's own model-zoo table marks Turbo's fine-tunability as "N/A," and outputs vary less across runs of the same prompt than the base model.
- Image editing is a separate Z-Image-Edit variant — Turbo itself is built for text-to-image.
- Turbo runs CFG-free with guidance scale fixed at 0, so negative prompts have less control than on the base Z-Image model.
Using it in Nomi
On the generation canvas, add a image card, pick “Z-Image Turbo” 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.
