Write a prompt
Describe the scene, the subject and the light. Specific prompts give better pictures than vague ones.
Generate Images with AI Online Free
Describe a picture and Z-Image-Turbo draws it on your own machine. The model is fetched once from Hugging Face and cached in your browser, so the 73hi server never carries the model and your prompt is never uploaded.
Graphics card
GPU not checked yet. Checking does not download model files.
Choose another GPU on Windows
Chrome on Windows uses the GPU assigned to the browser. This page cannot list all installed graphics cards or switch to a card that the browser does not expose.
- Open Windows Settings → System → Display → Graphics settings. Add the browser you use for this tool, such as chrome.exe or msedge.exe.
- Open the app's Options and check the GPU name shown under High performance. Select it if it is the card you want, then save. If Windows offers a specific GPU option, select your desired card there.
- For Chrome, you can also open
chrome://flags/#force-high-performance-gpu, enable the option if available, and restart Chrome. This requests the high-performance GPU; it does not select a card by name. - Fully close and restart the browser, return here and select Check GPU. Verify the reported GPU and f16 support before selecting Load model.
If both Windows options show the same unwanted card, choosing High performance alone does not select your other GPU. The browser must expose that card first. GPU names may be reported as vendor and architecture instead of the full model name.
Chrome's Windows GPU limitationsLoad the model first, then enter a prompt and generate.
The first Load model downloads about 5.5 GB from Hugging Face, so it can take several minutes and longer on a slow connection. Keep this tab open until the bar finishes. Files are cached when browser storage is available; selected local files avoid the download. A small GPU test runs before any model files are fetched. Your browser must expose a GPU with WebGPU f16 shader support. On computers with multiple GPUs, you may need to select a compatible GPU for the browser in your system graphics settings and fully restart the browser.
Your picture appears here
Model files
Every file name above is a direct link, so clicking one downloads just that file; the
.onnx_data companions must be fetched together with their graph for a manual load to work.
The INT4 ONNX build lives at
huggingface.co/webnn/Z-Image-Turbo.
Hugging Face allows cross origin requests, so this page pulls it straight from the Hub.
Three simple steps
How to Generate an AI Image Online
Draw a picture from words without an account, a queue or a server round trip.
Load the model
Select Load model and wait until the files are ready, then select Generate. Generate uses the loaded files and never downloads the model.
Save the picture
When the denoising steps finish, the image is drawn on the canvas and ready to download as a PNG.
How it works
A Diffusion Model, Entirely on Your Device
The whole pipeline runs in the page: the prompt is encoded, a latent is denoised for a handful of steps, and the VAE turns it into pixels.
| Stage | What it does | Where it runs |
|---|---|---|
| Tokenizer and text encoder | Turns your prompt into embeddings the model understands | Your browser |
| Transformer | Predicts the noise to remove at each step of the schedule | Your browser |
| Scheduler helper | Applies the flow matching step and produces the next latent | Your browser |
| VAE decoder | Converts the finished latent into a visible image | Your browser |
One time cost: the model files total about 5.5 GB. They come from Hugging Face, not from 73hi, and your browser caches them for the next visit.
Free AI image generator
Private, Uncensored by Quotas, No Watermark
Everything needed to make a picture without handing your prompt to a third party.
Nothing is uploaded
The prompt and the picture never leave your device. Only the model is downloaded.
GPU acceleration
With WebGPU the denoising loop runs on your graphics card, so a picture takes seconds rather than minutes.
Full control
Pick the resolution, the number of steps and the seed, and reproduce any picture exactly.
No quota
No signup, no daily limit, and no watermark on anything you generate.
Questions and answers
AI Image Generator FAQ
Where does the model come from?
The page downloads the INT4 ONNX build of Z-Image-Turbo directly from Hugging Face, which serves it with permissive cross origin headers. The 73hi server never stores or forwards the model.
Why is the download so large?
The text encoder and the transformer together are about 5.5 GB even in 4 bit quantisation. They are cached in your browser after the first run.
Do I need a GPU?
Yes. The transformer weights alone are 3.44 GB, and the WebAssembly backend is built for wasm32, whose entire memory is capped at 4 GB, so the model simply does not fit there. WebGPU streams the weights to your graphics card instead, which is the only way this model runs in a browser. Use Chrome or Edge 130 or newer.
Is my prompt uploaded?
No. The prompt is encoded and generated entirely on your device. Only the model and the small tokenizer are downloaded.
Why does 1024 x 1024 fail while 512 x 512 works?
The weights are the same at every resolution, but the attention maps inside the transformer grow with the square of the number of image tokens. A 512 x 512 image works on 64 x 64 latents, which is 4,096 tokens, while 1024 x 1024 works on 128 x 128 latents, which is 16,384 tokens and sixteen times the intermediate data. That is graphics memory, not WebAssembly: the runtime already asks the adapter for the largest limits it offers. If 1024 fails on your card, use 768 first, and the tool frees the text encoder before the denoise loop at 768 and above to leave as much memory as possible for the image itself.
How does the negative prompt work?
Leave it empty and nothing changes. Fill it in and set Guidance above 1 to run classifier free guidance: the model is evaluated against both prompts each step and the results are combined, which steers the picture away from whatever you listed. Guidance 1 means no steering. Because Z-Image-Turbo is a distilled few-step model, keep the value low, around 1.2 to 2, since higher values tend to oversaturate the image. Note that guidance doubles the number of transformer passes, so a picture takes roughly twice as long.
Can I use files I already downloaded?
Yes. Pick the model files with the file chooser and the tool uses the local copies instead of downloading them again.