# Z-Image-Turbo Online: Official Demo and Local Access Guide > Published 2026-04-01, updated 2026-09-06 ยท https://www.promptzone.com/florence_herrera/generation-z-ai-image-tool-breaks-speed-records-3n7o --- title: "Z-Image-Turbo Online: Official Demo and Local Access Guide" published: true description: "Try Z-Image-Turbo through its official online demo, understand the controls, and reproduce a first generation with the published Python pipeline." tags: ai,imagegeneration,comfyui,tutorial --- To use Z-Image-Turbo online, open the Tongyi-MAI team's official Hugging Face Space, enter a prompt, choose image dimensions, and select **Generate**. Save the result with its **Seed Used** value to repeat a trial with **Random Seed** disabled; the project also publishes weights for local inference. [Official demo][s4], [Demo controls][s3], [Project README][s1] ## What are the key facts about Z-Image-Turbo online access? | Field | Verified information | | --- | --- | | Developer | Tongyi-MAI's Z-Image team. [Project README][s1] | | Released | November 26, 2025, according to the project's release log. [Release log][s1] | | Type | Distilled text-to-image generation model. [Model card][s2] | | Size or parameters | 6 billion model parameters, as reported by the project. [Model card][s2] | | License and access | Apache-2.0 weights; official browser demo and downloadable inference pipeline. [Model card][s2] | | Where it runs | Hosted in the official demo, or on your own compatible inference hardware. The published Python example uses CUDA. [Demo implementation][s3], [README][s1] | ## Which controls does the official Z-Image-Turbo demo provide? The model card emphasizes photographic imagery, instruction following, and English and Chinese text rendering. Treat those as the developers' stated areas of strength. A useful trial should include the kind of prompt you actually expect to use. [Model card][s2] The demo defines **Prompt**, **Resolution Category**, **Width x Height (Ratio)**, **Seed**, **Random Seed**, **Generate**, and **Seed Used** controls. Its **Steps** slider displays eight and is not interactive; the source passes that value plus one to inference. [Demo source][s3] For a first task, try a simple still life with one short visible label. For example, ask for a yellow notebook beside a green pencil, with the word NOTES centered on the notebook. Specify the camera angle and background separately. Before generating, write down what would count as success: correct objects, correct colors, readable lettering, and the requested placement. This suggested rubric gives your experiment a purpose beyond picking the most attractive thumbnail. Keep the original prompt when you save your result. If you later compare another host or a local run, you will have a clear starting brief instead of a remembered approximation. Record the displayed model name as well. ## What should you know about Z-Image-Turbo speed and settings? The official performance claim of sub-second inference refers to enterprise H800 hardware. It is not a promise about browser response time, every hosting provider, or a consumer GPU. The model card separately discusses consumer-device memory fit. [Performance context][s2] Measure your own workflow from submission to a usable result if waiting time matters. Keep that measurement separate from any quoted model-only inference result. Label your notes with the host and settings so you do not mistake different conditions for a model comparison. Turbo uses a distilled sampling configuration. The published Diffusers example sets `guidance_scale=0.0` and `num_inference_steps=9`, explaining that this produces eight transformer evaluations. Follow those settings when establishing a baseline. [Official example][s1] The Z-Image foundation checkpoint uses different guidance and sampling recommendations. For that separate workflow, see the [Z-Image Base guide](/pietro_lefevre/sortie-z-image-base-a-new-ai-model-for-image-generation-2plm). Do not carry a Base preset into a Turbo experiment without checking the variant. Before using the online demo for a deadline, inspect its current access conditions and time a representative request. Save the host, submission time, image dimensions, and settings with your result. ## How do you use the Z-Image-Turbo demo and local Python pipeline? Open the official repository or model card and follow its Z-Image-Turbo online-demo link. Confirm that the Space belongs to **Tongyi-MAI** and identifies the Turbo variant before entering your prompt. The project links this Space directly. [Official access link][s1] Enter your prompt in the **Prompt** field. Choose a resolution category and an available width-and-height option, then select **Generate**. For the first attempt, leave the remaining controls at their defaults. These controls are defined in the published application source. [Demo controls][s3] Afterward, save the image and note **Seed Used**. To explore a wording change while retaining the seed, turn off **Random Seed** and enter the recorded value in **Seed**. Keep the image dimensions unchanged for that comparison. [Seed controls][s3] Make one edit to your prompt, such as moving the pencil from beside the notebook to above it. Review both images against the same requirements. This is a suggested controlled trial; it does not promise identical outputs across different software versions or hosts. If you want your own execution environment, the project's Python route uses Hugging Face Diffusers. Install PyTorch for your hardware and a Diffusers build that includes `ZImagePipeline`; the README documents installation from the official Diffusers repository. [Setup instructions][s1] ```python import torch from diffusers import ZImagePipeline pipe = ZImagePipeline.from_pretrained( "Tongyi-MAI/Z-Image-Turbo", torch_dtype=torch.bfloat16, low_cpu_mem_usage=False, ).to("cuda") result = pipe( prompt="A yellow notebook marked NOTES beside a green pencil.", width=1024, height=1024, num_inference_steps=9, guidance_scale=0.0, generator=torch.Generator("cuda").manual_seed(42), ) result.images[0].save("notebook.png") ``` This adapts the official inference example with an original prompt. The dimensions are example settings, not a claim about the model's maximum resolution. The example requires compatible CUDA hardware and enough memory for the pipeline. [Inference example][s1] Start by checking that the unmodified pipeline loads and writes an image. Only then introduce your application logic, batch processing, or performance changes. Keep the first successful environment's package versions alongside the prompt and output. For a visual local setup, continue with the sibling [Z-Image ComfyUI workflow](/kareem_kim/z-image-comfyui-a-new-tool-for-stable-diffusion-workflows-2n3e). Use that page for component placement and workflow selection instead of guessing from another model's directory layout. ## How does Z-Image-Turbo differ from the Z-Image foundation model? | Option | Documented emphasis | Suggested reason to evaluate it | | --- | --- | --- | | Z-Image-Turbo | Distilled generation with a low evaluation count. [Model card][s2] | You want to try the official demo and establish a compact generation workflow. | | Z-Image foundation model | Generation with guidance, negative prompts, and a focus on diversity and downstream development. [README][s1] | Your experiment needs those controls or a foundation checkpoint. | For broader workflow context, the [ComfyUI complete guide](/tomas_novak/comfyui-2026-the-complete-guide-to-power-user-ai-image-generation-1g17) explains the application surrounding a model. Choose the variant first, then the interface and hosting arrangement that suit the work. ## What are common Z-Image-Turbo online demo questions? ### Can I use Z-Image-Turbo without installing it? The Z-Image project links an official Tongyi-MAI Hugging Face Space for browser generation. Open that demo to try Z-Image-Turbo and check the access conditions shown by the host. [Official demo][s4], [Project links][s1] ### Are Z-Image-Turbo weights available? Tongyi-MAI publishes Z-Image-Turbo weights under Apache-2.0 on Hugging Face. The official repository also provides a Diffusers example for running the checkpoint on compatible hardware. [Model card][s2], [Inference example][s1] ### Why does the Z-Image-Turbo Python example use nine steps? The official Z-Image-Turbo Diffusers example uses `num_inference_steps=9`, which the project explains produces eight transformer forward evaluations. It pairs this with `guidance_scale=0.0`; use both settings for the documented Turbo baseline. [README example][s1] ### How do I keep the same seed in the Z-Image-Turbo demo? Copy the result's **Seed Used** value into **Seed** and turn off **Random Seed**. Keep the dimensions and other settings unchanged when comparing prompt wording in the same demo environment. [Demo source][s3] ## Sources - [Z-Image official repository and inference examples][s1] - [Z-Image-Turbo official model card][s2] - [Official demo application source][s3] - [Official Z-Image-Turbo online demo][s4] [s1]: https://github.com/Tongyi-MAI/Z-Image [s2]: https://huggingface.co/Tongyi-MAI/Z-Image-Turbo [s3]: https://huggingface.co/spaces/Tongyi-MAI/Z-Image-Turbo/blob/main/app.py [s4]: https://huggingface.co/spaces/Tongyi-MAI/Z-Image-Turbo ## Related guides on PromptZone - [Best SDXL Models in 2026](/tara_suzuki/best-sdxl-models-in-2026-realistic-anime-and-all-purpose-checkpoints-116) - [ComfyUI 2026: The Complete Guide](/tomas_novak/comfyui-2026-the-complete-guide-to-power-user-ai-image-generation-1g17) - [AI Model Releases Timeline](/ai-model-releases)