Skip to main content
Endpoint: POST https://fal.run/fal-ai/z-image/turbo/controlnet Endpoint ID: fal-ai/z-image/turbo/controlnet

Try it in the Playground

Run this model interactively with your own prompts.

​
Quick Start

import fal_client

def on_queue_update(update):
    if isinstance(update, fal_client.InProgress):
        for log in update.logs:
           print(log["message"])

result = fal_client.subscribe(
    "fal-ai/z-image/turbo/controlnet",
    arguments={
        "prompt": "A single leopard, its spotted golden coat detailed with black rosettes, cautiously peeks its head through dense green foliage. The leopard’s eyes are alert and focused forward, ears perked, whiskers slightly visible. The bushes consist of thick, leafy shrubs with varying shades of green, some leaves partially obscuring the leopard’s muzzle and forehead. Soft natural daylight filters through the canopy above, casting dappled shadows across the animal’s fur and surrounding leaves. The composition is a medium close-up, centered on the leopard’s head emerging from the undergrowth, with shallow depth of field blurring the background vegetation.",
        "image_url": "https://storage.googleapis.com/falserverless/example_inputs/z-image-turbo-controlnet-input.jpg"
    },
    with_logs=True,
    on_queue_update=on_queue_update,
)
print(result)

​
Input Schema

​
prompt
string
required
The prompt to generate an image from.
​
image_size
ImageSize | Enum
default:"auto"
The size of the generated image. Default value: autoPossible values: square_hd, square, portrait_4_3, portrait_16_9, landscape_4_3, landscape_16_9, auto
​
num_inference_steps
integer
default:"8"
The number of inference steps to perform. Default value: 8Range: 1 to 8
​
seed
integer
The same seed and the same prompt given to the same version of the model will output the same image every time.
​
sync_mode
boolean
default:"false"
If True, the media will be returned as a data URI and the output data won’t be available in the request history.
​
num_images
integer
default:"1"
The number of images to generate. Default value: 1Range: 1 to 4
​
enable_safety_checker
boolean
default:"true"
If set to true, the safety checker will be enabled. Default value: true
​
output_format
OutputFormatEnum
default:"png"
The format of the generated image. Default value: "png"Possible values: jpeg, png, webp
​
acceleration
AccelerationEnum
default:"regular"
The acceleration level to use. Default value: "regular"Possible values: none, regular, high
​
enable_prompt_expansion
boolean
default:"false"
Whether to enable prompt expansion. Note: this will increase the price by 0.0025 credits per request.
​
image_url
string
required
URL of Image for ControlNet generation.
​
control_scale
float
default:"0.75"
The scale of the controlnet conditioning. Default value: 0.75Range: 0 to 1
​
control_start
float
default:"0"
The start of the controlnet conditioning.Range: 0 to 1
​
control_end
float
default:"0.8"
The end of the controlnet conditioning. Default value: 0.8Range: 0 to 1
​
preprocess
Enum
default:"none"
What kind of preprocessing to apply to the image, if any. Default value: nonePossible values: none, canny, depth, pose

​
Output Schema

​
images
list<ImageFile>
required
The generated image files info.
​
timings
Timings
required
The timings of the generation process.
​
seed
integer
required
Seed of the generated Image. It will be the same value of the one passed in the input or the randomly generated that was used in case none was passed.
​
has_nsfw_concepts
list<boolean>
required
Whether the generated images contain NSFW concepts.
​
prompt
string
required
The prompt used for generating the image.

​
Input Example

{
  "prompt": "A single leopard, its spotted golden coat detailed with black rosettes, cautiously peeks its head through dense green foliage. The leopard’s eyes are alert and focused forward, ears perked, whiskers slightly visible. The bushes consist of thick, leafy shrubs with varying shades of green, some leaves partially obscuring the leopard’s muzzle and forehead. Soft natural daylight filters through the canopy above, casting dappled shadows across the animal’s fur and surrounding leaves. The composition is a medium close-up, centered on the leopard’s head emerging from the undergrowth, with shallow depth of field blurring the background vegetation.",
  "image_size": "auto",
  "num_inference_steps": 8,
  "sync_mode": false,
  "num_images": 1,
  "enable_safety_checker": true,
  "output_format": "png",
  "acceleration": "regular",
  "enable_prompt_expansion": false,
  "image_url": "https://storage.googleapis.com/falserverless/example_inputs/z-image-turbo-controlnet-input.jpg",
  "control_scale": 0.75,
  "control_start": 0,
  "control_end": 0.8,
  "preprocess": "none"
}

​
Output Example

{
  "images": [
    {
      "content_type": "image/png",
      "height": 1024,
      "url": "https://storage.googleapis.com/falserverless/example_outputs/z-image-turbo-controlnet-output.jpg",
      "width": 1536
    }
  ],
  "prompt": ""
}
Tongyi-MAI’s Z-Image Turbo delivers ControlNet-guided image generation at $0.0065 per megapixel through a 6-billion parameter architecture. This model trades raw parameter count for specialized control mechanisms, canny edge detection, depth mapping, and pose guidance, that preserve structural fidelity during image-to-image transformations. Built for designers and developers who need precise spatial control without the inference overhead of larger diffusion models. Use Cases: Product visualization with reference geometry | Character pose transfer workflows | Architectural rendering from depth maps

​
Performance

Z-Image Turbo operates at roughly 3-5x more cost-effective rates than traditional ControlNet implementations by optimizing the 6B parameter base for rapid inference. At $0.0065 per megapixel, you’re running 153 megapixels per dollar, ideal for batch processing workflows where structural guidance matters more than photorealistic perfection.
MetricResultContext
Model Size6 billion parametersOptimized for inference speed vs 70B+ alternatives
Inference Steps1-8 configurableDefault 8 steps balances quality and latency
Cost per Megapixel$0.0065153 megapixels per $1.00 on fal
Control Methods4 preprocessing modesNone, canny edge, depth map, pose detection
Batch GenerationUp to 4 images per requestParallel generation with shared control input
Related EndpointsStandard image-to-image, LoRA variantsControlNet vs direct transformation vs custom training

​
Structural Control Without Compromise

Z-Image Turbo routes your prompt through three parallel conditioning pathways: text embedding, reference image structure, and optional preprocessing filters. Unlike pure text-to-image models that hallucinate spatial relationships, this architecture extracts edge maps, depth channels, or skeletal poses from your input, then enforces those constraints during diffusion. What this means for you:
  • Configurable control strength (0-1 scale): Dial conditioning intensity from 0.9 for strict adherence to 0.3 for loose interpretation, critical when your reference image has good composition but needs significant style deviation
  • Temporal control windowing: Apply ControlNet guidance only during steps 0-40% of generation (configurable start/end), letting early diffusion lock structure while late steps refine aesthetics
  • Four preprocessing modes: Feed raw images directly or auto-extract canny edges (sharp boundaries), depth maps (spatial layering), or pose skeletons (human/character positioning) without external tools
  • Multi-format output with safety: Generate 1-4 variants simultaneously in JPEG, PNG, or WebP with optional built-in safety filtering, batch testing style variations while maintaining structural consistency

​
Technical Specifications

SpecDetails
ArchitectureZ-Image Turbo 6B
Input FormatsText prompt + reference image URL (JPEG, PNG, WebP, GIF, AVIF)
Output FormatsJPEG, PNG, WebP with configurable dimensions
Preprocessing OptionsNone, Canny edge detection, Depth estimation, Pose detection
Control ParametersScale (0-1), temporal start/end windowing, inference steps (1-8)
LicenseCommercial use permitted
API Documentation | Quickstart Guide | Enterprise Pricing

​
How It Stacks Up

**Z-Image Turbo Standard (0.0065/MP)∗∗–The[ControlNetvariant](https://fal.ai/models/fal−ai/z−image/turbo/controlnet)addsstructuralguidancepreprocessingfor0.0065/MP)** – The [ControlNet variant](https://fal.ai/models/fal-ai/z-image/turbo/controlnet) adds structural guidance preprocessing for 0.0065 per megapixel, same base cost. Standard image-to-image prioritizes direct style transfer without intermediate edge/depth extraction, ideal for texture swaps and color grading where spatial relationships already match your target. ControlNet trades processing simplicity for precise geometric control when your reference structure needs enforcement. FASHN Virtual Try-On V1.5 – Z-Image Turbo ControlNet offers general-purpose structural conditioning across edge, depth, and pose modalities for diverse creative workflows. FASHN specializes in garment-to-body fitting with proprietary try-on algorithms optimized for fashion e-commerce, trading generality for domain-specific accuracy in clothing visualization.

​
Limitations

  • image_size restricted to: square_hd, square, portrait_4_3, portrait_16_9, landscape_4_3, landscape_16_9, auto
  • num_inference_steps range: 1 to 8
  • num_images range: 1 to 4
  • output_format restricted to: jpeg, png, webp
  • acceleration restricted to: none, regular, high
  • control_scale range: 0 to 1
  • control_start range: 0 to 1
  • control_end range: 0 to 1
  • preprocess restricted to: none, canny, depth, pose
  • Content moderation via safety checker