README.md

metadata

title: Z-Image / Z-Image-Turbo
emoji: ⚡
colorFrom: indigo
colorTo: pink
sdk: gradio
sdk_version: 6.20.0
app_file: app.py
pinned: false
license: apache-2.0
short_description: Text-to-image generation

Z-Image / Z-Image-Turbo Gradio demo

A single-file Gradio app for Tongyi-MAI's Z-Image models:

• Tongyi-MAI/Z-Image — full, undistilled base model. Supports CFG and negative prompts, 28–50 steps.
• Tongyi-MAI/Z-Image-Turbo — distilled for speed, 8 steps, no CFG.

Both load through the same diffusers ZImagePipeline class, so app.py works for either model — which one loads is controlled by the MODEL_ID constant at the top of the file.

Deploying

1. Create a new Space at huggingface.co/new-space, SDK = Gradio.
2. Upload app.py, requirements.txt, and this README.md.
3. Under Settings → Hardware, select a GPU. For a free public demo, choose ZeroGPU (requires a PRO / Team / Enterprise account to enable it, but is free to run once enabled); otherwise pick a dedicated GPU tier with at least ~16 GB VRAM.
4. To serve the base model instead of Turbo, either edit the MODEL_ID line in app.py to "Tongyi-MAI/Z-Image", or add a Space Variable named MODEL_ID with that value and leave the code untouched. Deploying the same repo twice with different MODEL_IDs (mirroring the two official demo Spaces) is the simplest way to offer both.

Notes

• low_cpu_mem_usage=False and installing diffusers from Git (not PyPI) are both required per the model cards as of this writing.
• The Turbo model disables the steps and CFG controls in the UI since deviating from its distilled 8-step schedule tends to hurt quality; the base model leaves them adjustable.
• ZeroGPU currently backs onto H200 GPUs with ~70GB VRAM per workload, so if you'd rather have one Space that lets people switch between both models live (instead of deploying two Spaces), that's very feasible memory-wise — ask and this file can be adapted to load both and add a model-select dropdown.
• This is a minimal starting point. The official Z-Image-Turbo Space additionally layers on prompt enhancement, an NSFW safety checker, and ahead-of-time (AoTI) compilation for extra speed, none of which are included here to keep things simple to read and deploy.