How long does one image take?

On Orin NX 16GB with 4 sampling steps, a 512px text-to-image takes about 100 s and 384px about 73 s. Image-to-image takes longer with more steps and edit strength, 1-2 minutes per image.

Does it work offline, and does data leave the device?

After deployment it runs fully offline; the first deploy needs internet access to download weights and inference engines. Prompts, reference images and results stay on the device, travelling only between your app and it.

Can several people use it at once?

The service handles one request at a time and queues the rest, to stay within GPU memory. Concurrent generation for many users needs several devices.

How do I connect it to my own application?

Call the HTTP API on port 8000. Send a prompt as JSON, or upload a reference image as a form for image-to-image. The response carries elapsed time and an image URL to fetch the PNG; set the client timeout to 180 s.

Downloads are slow or interrupted. What now?

The deploy script falls back to a mirror when Hugging Face is blocked. After an interruption, just redeploy; partial files are discarded and only missing ones are fetched. If both are blocked, place the model in the model directory by hand and redeploy.

Can I change resolution or step count?

Resolution is chosen at deployment, 512px or 384px, and each additional resolution downloads about 12GB of inference engine. Steps and image-to-image strength are set per request; more steps take longer.

What should we check on model licensing for a commercial product?

The Z-Image-Turbo weights keep their publisher's licence terms, and the deployment code in the upstream GitHub repository is used under its licence. Before shipping a closed product, confirm the weight licence and any limits on generated content suit your product.

Z-Image-Turbo on Jetson

Run Z-Image-Turbo (6B) locally on Jetson Orin NX — text-to-image and img2img via HTTP API, fully offline.

SenseCraftIntermediate15minOn-Device LMSource Code

Reusable Site Templates

Places where images must not leave the device and there is no reliable internet, but waiting a minute or two per image is acceptable.

Privacy-sensitive image workflows — generate or edit images inside healthcare, legal or enterprise networks; images and prompts stay on the device.
Offline interactive installations — photo booths and gallery pieces that draw what a visitor describes, with no cloud service.
On-device image generation in your app — a web app, desktop tool or automation calls the HTTP API and gets a PNG back.

Not a fit: real-time generation or many users generating at once. Each image takes 1-2 minutes, and one request runs at a time.

How it works

Z-Image-Turbo device composition: a reComputer J40 / J50 Series host returns generated images to apps over HTTP

One Jetson host is all it takes. The model and generation service sit in one Docker container and open an HTTP API on the LAN; your app sends a prompt or reference image and gets a PNG back.

A 6-billion-parameter image model means the host needs a large-memory module.

Host	Module	Notes
reComputer J40 Series (J4012)	Orin NX 16GB	The published generation times are for this module
reComputer J50 Series	AGX Orin 64GB	Same deployment flow

The host needs JetPack 6, the NVIDIA container runtime and SSH access.

The first deploy downloads the model weights and the inference engine for the chosen resolution, so check disk space first.

Weights are about 20GB, and each resolution's inference engine adds about 12GB.
Downloads resume: redeploy after an interruption and only missing files are fetched; if Hugging Face is blocked it falls back to a mirror.
One resolution: 512px or 384px, chosen at deployment.

How well it works

Scope: whole-image generation time from sending the prompt to receiving the PNG, 4 sampling steps, one request.

What the site gets	Typical	Device
512px text-to-image	about 100 s	reComputer J40 Series (J4012, Orin NX 16GB)
384px text-to-image	about 73 s	Same
Requests handled at once	1, the rest queue	Same

Image-to-image takes a reference image plus a prompt; time grows with steps and edit strength, landing at 1-2 minutes per image. Set the HTTP client timeout to 180 s.

What you get out of it

Once deployed, these are the interfaces.

Interface	What comes out	Notes
HTTP 8000 health check	Whether the service is ready	Returns success when ready
HTTP 8000 JSON generation	Text-to-image or image-to-image; returns mode, resolution, elapsed time and image URL	For programmatic calls
HTTP 8000 form generation	Same, with the reference image as a file upload	For sending a local image directly
HTTP 8000 image files	The generated PNG	Also saved to the output directory on the device

The API is served on the LAN and runs one request at a time; a short client timeout disconnects before the image is done.

How to deploy

Check the setup and its product list, then follow the deployment guide.

Deployment Options

Deploy Image Generation Service

Deploy Z-Image-Turbo HTTP API to Jetson Orin NX via SSH

2 steps · ~15min

Open the deployment guide →

Jetson Device

recomputer_j40

recomputer_j50

FAQ

The minimum is a reComputer J40 Series model with the Orin NX 16GB module, such as J4012; the reComputer J50 Series (AGX Orin 64GB) also works. The host needs JetPack 6 and disk for about 20GB of weights plus about 12GB of inference engine per resolution.

Didn't find your answer?

We Are Glad to Be Your Hardware Partner !