Which LoRAs work here?

Files trained for Z-Image Turbo, of which Hugging Face lists hundreds. Pass the direct download link to the .safetensors file. LoRAs made for FLUX, SDXL or Qwen-Image load without an error and do little, so run a with-and-without test before you tune anything.

Can I use uncensored LoRAs?

Yes. The weights are not reviewed and nothing filters the output. With a LoRA trained for unfiltered content loaded, our strongest content probe rendered as written, and even without a LoRA the base model did not hold back in our tests. Rights and consent for everything you generate stay with you.

Why did my LoRA request fail?

Usually the link needs a login, since gated Hugging Face repositories answer with HTTP 401, or it points to a web page instead of the file. Use a public URL that downloads the .safetensors file itself. Failed requests are not charged.

How big are the images?

Text-to-image returns about one megapixel: 1024 x 1024 at 1:1, 1344 x 768 at 16:9 and 1536 x 640 at 21:9. Editing keeps the source's aspect ratio and came back with a long side of 1024 pixels in our tests. The price is the same at every size.

What strength should I use when editing?

Start at the default 0.8, which restyles while keeping layout and pose. At 0.6 the change is subtle; near 1 the source is only a loose guide. Keep the seed fixed and move in steps of 0.1, so strength is the only thing changing between two results.

Should I use this, FLUX.1 Dev LoRA or Qwen Image 2512 LoRA?

A LoRA only works on the base it was trained for, so the file you have usually decides. When you get to choose: this family for speed, volume and editing with a strength control; FLUX.1 Dev LoRA when the look you want already exists among its more than a thousand community LoRAs, for new images only; Qwen Image 2512 LoRA for frames up to 1536 pixels and edits from one to three pictures.

Z-Image Turbo LoRA API: Uncensored NSFW Images, Your LoRAs

alibaba/z-image-turbo-lora/text-to-image

Z-Image Turbo LoRA text-to-image runs Alibaba's 8-step Z-Image Turbo with up to three of your own LoRAs, each with a strength from 0 to 4. It returns about one megapixel in nine aspect ratios as JPEG, PNG or WebP.

From $0.012/image

Run model View API

Input

Text to image

Prompt475/5000

Aspect ratio

Style LoRAs1/3

Realism

Output format

JPEGPNGWebP

Seed

Safety checkerOff

SpicyAPI does not run a safety checker on this model. Your prompt and media go to the model as written, and the model itself decides what comes back.

Checking your account before enabling submission and pricing.

Output

Ready for a new run

Example result

Animate

Edit

Split into layers

{
  "model": "alibaba/z-image-turbo-lora/text-to-image",
  "input": {
    "aspect_ratio": "3:4",
    "loras": [
      {
        "path": "https://huggingface.co/suayptalha/Z-Image-Turbo-Realism-LoRA/resolve/main/pytorch_lora_weights.safetensors",
        "scale": 0.7
      }
    ],
    "output_format": "jpeg",
    "seed": 42,
    "prompt": "Realism, cinematic film still, a glamorous woman in her early thirties with glossy dark waves and red lipstick, wearing a burgundy satin slip dress with one strap slipping off her shoulder, leaning on a black grand piano in a smoky 1950s jazz lounge at night, magenta and teal gel lights cutting through the haze, warm amber rim light on her bare shoulders, glossy reflections on the piano lid, a knowing half-smile toward the camera, shallow depth of field, 35mm film grain."
  }
}

Introduction

Z-Image Turbo LoRA Text to image makes one image from a written prompt on Alibaba's Z-Image Turbo, a 6-billion-parameter model distilled to eight sampling steps, with up to three of your own LoRA files loaded. Pick any of nine aspect ratios at about one megapixel and get back JPEG, PNG or WebP at one flat price. The platform adds no content filter to the prompt, the weights or the image.

Choose a mode by starting material

What you have	Mode
a written description	Z-Image Turbo LoRA Text to image
a picture to restyle	Z-Image Turbo LoRA Image to image

Key features & highlights

Eight steps, seconds per image

Z-Image Turbo is distilled to eight sampling steps. In our runs the model took roughly three to ten seconds per image, which makes it practical for testing LoRA scales and prompts in quick loops.

A large library of community LoRAs

Z-Image Turbo is one of the most popular bases for community fine-tunes, with hundreds of LoRAs published on Hugging Face: realism, painting, pixel art, characters and concepts other hosts refuse. Pass the direct file link and a scale; nothing is uploaded to us in advance.

Uncensored with custom weights

With a matching LoRA loaded, our strongest content probe came back as written, and the base model did not hold back without one. The weights are never reviewed and the output is never filtered.

Request parameters

Request parameters for Z-Image Turbo LoRA · Text to Image
Parameter	Required	Description
`prompt`	Yes	Put the LoRA's trigger word first, then the subject, the setting and the light. Plain descriptive sentences work better than keyword lists. at least 1 characters · at most 5000 characters
`aspect_ratio`	No	Nine options from 21:9 to 9:21, all about one megapixel and all the same price. `1:1` · `16:9` · `9:16` · `4:3` · `3:4` · `3:2` · `2:3` · `21:9` · `9:21` · Default: `1:1`
`loras`	No	Up to three {path, scale} entries trained for Z-Image Turbo; scale 0 to 4, 1 is neutral. Use a public direct link to the .safetensors file. Default: `[]` at most 3 items
`output_format`	No	JPEG by default, PNG or WebP if you need them. Does not change the price. `jpeg` · `png` · `webp` · Default: `jpeg`
`seed`	No	Keeps a set of images comparable while you change one LoRA scale at a time. `≥ 0` · `≤ 2147483647`

Pricing

$0.012 / image

Billed per image at one flat price for every aspect ratio. Loading LoRAs adds no separate charge.

Use cases

Style batches at a flat price

Every image costs the same whatever the ratio, so a style LoRA plus a list of prompts becomes a consistent set of illustrations, posters or thumbnails with a predictable bill. Fix the seed per layout and swap only the subject line.

Character sheets

Keep a character LoRA and a seed fixed and change only the outfit, place and light between prompts. The face and build stay recognisable across the whole set, which is what a sheet needs.

Uncensored stills for your platform

Load the LoRA your audience expects and write the image directly. There is no content flag and no second review of the result, so what your users ask for is what they get back.

Explore key use-case tutorials

Prompt guide & pro tips

Trigger word first

Most Z-Image Turbo LoRAs are trained with a trigger such as "Realism" or "Pixel art style". Put it at the very start, then describe the scene in plain sentences: who is in the frame, what they wear, where they are and what time of day it is.

Describe light, not adjectives

Name the light source, its direction and its colour: rain on the window, a warm lamp on the face, neon behind. Concrete lighting lands better than words like "cinematic", and it is the fastest way to make a realism LoRA look like a photograph.

Change one scale at a time

When stacking LoRAs, fix the seed and move one scale by about 0.2 at a time. Two style LoRAs at full strength usually fight; start the second one around 0.5 and raise it until its look appears without swallowing the first.

Draft in the shape you will publish

Pick the aspect ratio before you iterate. Composition changes between 1:1 and 21:9, and since every ratio costs the same there is no reason to draft square and crop later.

API quick start

# Generate the idempotency key once; reuse it when retrying after a lost response.
export SPICY_IDEMPOTENCY_KEY="${SPICY_IDEMPOTENCY_KEY:-$(uuidgen)}"
curl -X POST https://api.spicyapi.ai/api/v1/jobs/createTask \
  -H "Authorization: Bearer $SPICY_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: $SPICY_IDEMPOTENCY_KEY" \
  -d '{
  "model": "alibaba/z-image-turbo-lora/text-to-image",
  "input": {
    "aspect_ratio": "1:1",
    "output_format": "jpeg",
    "prompt": "A ceramic teapot on a linen cloth, soft window light, muted palette"
  }
}'

# Copy data.taskId from the response into TASK_ID, then poll the status.
TASK_ID="TASK_ID_FROM_RESPONSE"
curl "https://api.spicyapi.ai/api/v1/jobs/recordInfo?taskId=$TASK_ID" \
  -H "Authorization: Bearer $SPICY_API_KEY"

See the API reference for authentication, the complete schema, polling, and error handling.

Related models and next tasks

Missing the model you need?

Tell us what to add and where you would use it. Requests go to the people who decide what gets onboarded next.

Test a LoRA in seconds, then run the batch at one flat price

Z-Image Turbo LoRA loads up to three of your own LoRA files onto Alibaba's Z-Image Turbo, a 6-billion-parameter model distilled to eight sampling steps. In our runs an image came back in about three to ten seconds, and every image costs the same at any aspect ratio and on both endpoints, with nothing added for the LoRAs. That makes it the quickest place to find out what a Z-Image Turbo LoRA really does, and a predictable place to run it at volume. There is no platform filter on the prompt, the weights or the image.

Open the playground

Test a LoRA in seconds, then run the batch at one flat price

LoRAs per request: Up to 3, each {path, scale}; scale 0 to 4, and 1 is neutral
Speed: About three to ten seconds per image in our runs, from 8 sampling steps
Text to image: About one megapixel in nine ratios, from 21:9 to 9:21
Editing: One source, strength 0 to 1 (default 0.8); keeps the source's shape, long side 1024 px in our tests
Price: One flat price per image at every size and on both endpoints; LoRAs add nothing
Content filter: None; with a LoRA trained for it, our strongest content probe rendered as written

Fast loop: A LoRA question answered before you lose the threadMost of what you need to know about a new LoRA comes from two images: the same prompt and seed, once with the file and once without. At a few seconds each, that pair, and the scale ladder that follows it, fits into one short sitting instead of an afternoon.
Flat price: The bill is a count of imagesAspect ratio, output format and the number of LoRAs loaded do not move the price, and the edit endpoint costs the same as text-to-image. Budget a batch by counting its images, and draft in the ratio you will publish instead of drafting square and cropping later.
Limits: What it will not doOne image per request at about one megapixel. Editing has no mask, redraws the whole frame and comes back smaller: 1024 pixels on the long side in our tests. For bigger frames or edits from several pictures, Qwen Image 2512 LoRA goes up to 1536 pixels and takes up to three sources, with LoRAs trained for its own base.

Working with LoRAs on Z-Image Turbo LoRA

Let the LoRA carry the look

In our rain-window example the prompt only sets the scene and the light: a woman on a window ledge at night, a warm lamp on her face, rain on the glass. The photographic finish comes from a community Realism LoRA, with its trigger word at the start of the prompt. Describe the picture in plain sentences and leave the style to the file.

Pick the strength before you rewrite the prompt

On the edit endpoint, strength decides how much of the source survives. Our stage photo went through a classic painting LoRA at 0.75: the singer, the velvet dress and the microphone stayed where they were, and the whole frame turned to brushwork. The default 0.8 restyles and keeps the layout, 0.6 barely changes the picture, and values near 1 treat the source as a loose guide.

Check the file was trained for this base
LoRAs made for FLUX, SDXL or Qwen-Image load without an error and then do very little here. If the same prompt and seed look the same with and without the file, it was almost certainly trained for another model. Hugging Face lists hundreds of LoRAs built for Z-Image Turbo, so start there.
Stack gently
Each request takes up to three LoRAs with their own scale from 0 to 4. Two style LoRAs at full strength usually fight each other, so bring a second one in around 0.5 and raise it in small steps, seed fixed, until its look appears without swallowing the first.
Treat the link as part of the request
The file is fetched when the request runs, so a large LoRA adds time and a link that needs a login fails the request; gated Hugging Face repositories answer with HTTP 401. Use a public URL that downloads the .safetensors file itself, not the page that describes it. Failed requests are not charged.

Who uses Z-Image Turbo LoRA

Illustration and social teams
A whole series in one style, at a known cost
Load a style LoRA such as pixel art, fix the seed per layout and swap only the subject line. Every image in the set shares the look and the price, whatever mix of aspect ratios your channels need.
Portrait and character work
Painted portraits in one sitting
A classic painting LoRA turns a written portrait into an oil painting with brushwork, varnish and candlelight. Keep the subject, the seed and the LoRA fixed and change only the setting or the pose to build a matching gallery.
Concept artists and illustrators
Drawings turned into photographs
Send a drawing to the edit endpoint with a Realism LoRA and the strength at 0.8. Our cartoon singer came back as a photograph of a woman in a green velvet dress at a vintage microphone: the framing carried over and the medium changed.
Platforms offering community LoRAs
Your users' LoRAs, run as they are
Community LoRAs of every kind load with no review, and nothing filters the result. Because every image has the same price whatever the size and however many LoRAs are loaded, you know what a user's request costs before it runs.

From first test to finished batch with Z-Image Turbo LoRA

Find a LoRA trained for Z-Image Turbo
Read its model card for the trigger word and a suggested strength, and copy the direct download link to the .safetensors file.
Run it with and without the file
Same prompt, same seed, one run each. A style LoRA should be obvious at a glance; if nothing changes, go back to the first step.
Ladder the scale
Hold everything else and move the scale about 0.2 at a time from 1. Keep the value where the look lands without taking over the subject.
Draft in the ratio you will publish
Composition changes between 1:1 and 21:9, and every ratio costs the same, so there is nothing to gain from drafting square.
Run the batch, and restyle with the same array
Send the settled loras array with every prompt. For existing pictures, use the edit endpoint with the same array and set the strength first, starting at 0.8.

Z-Image Turbo LoRA prompts worth copying

A with-and-without pair

Run it once with a Realism LoRA and once with loras empty, on the same seed. The difference between the two frames is everything the file contributes.

Realism, a woman in a dark wool coat waiting at a tram stop at night, rain on the shelter glass, warm light from a kiosk on her face, city lights blurred behind her, 35mm film grain

A series line for a style LoRA

With a pixel-art LoRA, keep this sentence shape and swap only the subject and the place; the trigger always stays first.

Pixel art style. Two friends sharing an umbrella on a neon-lit street at night, puddles reflecting the signs, steam rising from a noodle stall, string lights overhead

A photo repainted, pose kept
For the edit endpoint with a painting LoRA at strength 0.75. Describe the whole result, not only the change, because the model redraws the entire frame.
```
A classical oil painting of the same man in a grey suit at a café table by the window, afternoon light, heavy brushwork, dark museum varnish
```

All Z-Image models5

Z-Image Base LoRAText to image · Image editing$0.014per image
Z-Image Spicy ProText to imageSpicy list price: $0.02$0.019per image
Z-Image SpicyText to imageSpicy list price: $0.013From $0.012per image
Z-ImageText to imageAlibaba official price: $0.015From $0.01per image
Z-Image Turbo LoRACurrentText to image · Image editing$0.012per image

Related comparisons

Z-Image Turbo LoRA vs FLUX.1 Dev LoRA

LoRA guide

Best NSFW LoRAs for Z-Image Turbo

Workbench