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

Input Schema

string
required
The prompt to generate an image from.
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
integer
default:"8"
The number of inference steps to perform. Default value: 8Range: 1 to 8
integer
The same seed and the same prompt given to the same version of the model will output the same image every time.
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.
integer
default:"1"
The number of images to generate. Default value: 1Range: 1 to 4
boolean
default:"true"
If set to true, the safety checker will be enabled. Default value: true
OutputFormatEnum
default:"png"
The format of the generated image. Default value: "png"Possible values: jpeg, png, webp
AccelerationEnum
default:"regular"
The acceleration level to use. Default value: "regular"Possible values: none, regular, high
boolean
default:"false"
Whether to enable prompt expansion. Note: this will increase the price by 0.0025 credits per request.
string
required
URL of Image for ControlNet generation.
float
default:"0.75"
The scale of the controlnet conditioning. Default value: 0.75Range: 0 to 1
float
default:"0"
The start of the controlnet conditioning.Range: 0 to 1
float
default:"0.8"
The end of the controlnet conditioning. Default value: 0.8Range: 0 to 1
Enum
default:"none"
What kind of preprocessing to apply to the image, if any. Default value: nonePossible values: none, canny, depth, pose

Output Schema

list<ImageFile>
required
The generated image files info.
Timings
required
The timings of the generation process.
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.
list<boolean>
required
Whether the generated images contain NSFW concepts.
string
required
The prompt used for generating the image.

Input Example

Output Example

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.

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

API Documentation | Quickstart Guide | Enterprise Pricing

How It Stacks Up

**Z-Image Turbo Standard (0.0065/MP)The[ControlNetvariant](https://fal.ai/models/falai/zimage/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