Python Runware SDK

The Python Runware SDK is used to run image inference with the Runware API, powered by the Runware inference platform. It can be used to generate images with text-to-image and image-to-image. It also allows the use of an existing gallery of models or selecting any model or LoRA from the CivitAI gallery. The API also supports upscaling, background removal, inpainting and outpainting, and a series of other ControlNet models.

Get API Access

To use the Python Runware SDK, you need to obtain an API key. Follow these steps to get API access:

1. Create a free account with Runware.
2. Once you have created an account, you will receive an API key and trial credits.

Important: Please keep your API key private and do not share it with anyone. Treat it as a sensitive credential.

Documentation

For detailed documentation and API reference, please visit the Runware Documentation or refer to the docs folder in the repository. The documentation provides comprehensive information about the available classes, methods, and parameters, along with code examples to help you get started with the Runware SDK Python.

Installation

To install the Python Runware SDK, use the following command:

pip install runware

Usage

Before using the Python Runware SDK, make sure to set your Runware API key in the environment variable RUNWARE_API_KEY. You can do this by creating a .env file in your project root and adding the following line:

RUNWARE_API_KEY = "your_api_key_here"

Generating Images

To generate images using the Runware API, you can use the imageInference method of the Runware class. Here's an example:

from runware import Runware, IImageInference

async def main() -> None:
    runware = Runware(api_key=RUNWARE_API_KEY)
    await runware.connect()

    request_image = IImageInference(
        positivePrompt="a beautiful sunset over the mountains",
        model="civitai:36520@76907",  
        numberResults=4,  
        negativePrompt="cloudy, rainy",
        height=512,  
        width=512, 
    )

    images = await runware.imageInference(requestImage=request_image)
    for image in images:
        print(f"Image URL: {image.imageURL}")

Enabling teaCache/deepCache/fbCache for faster inference

Some models support teaCache, deepCache, and fbCache for faster inference, with the trade-off of quality loss with more aggressive settings.

from runware import Runware, IImageInference, IAcceleratorOptions

async def main() -> None:
    runware = Runware(api_key=RUNWARE_API_KEY)
    await runware.connect()

    request_image = IImageInference(
        positivePrompt="a beautiful sunset over the mountains",
        model="civitai:943001@1055701", # using Shuttle v3 for this test, to showcase the power on 3rd party Flux finetunes.
        numberResults=1,
        negativePrompt="cloudy, rainy",
        height=1024,
        width=1024,
        acceleratorOptions=IAcceleratorOptions(
            teaCache=True,
            teaCacheDistance=0.6, # 0.6 is at the more moderate-to-extreme end, and 0.1 is at the more conservative end.
        ),
    )

    images = await runware.imageInference(requestImage=request_image)
    for image in images:
        print(f"Image URL: {image.imageURL}")

Using fbCache for enhanced performance

fbCache (First Block Cache) provides additional acceleration options for compatible models:

from runware import Runware, IImageInference, IAcceleratorOptions

async def main() -> None:
    runware = Runware(api_key=RUNWARE_API_KEY)
    await runware.connect()

    request_image = IImageInference(
        positivePrompt="a futuristic cityscape with flying cars",
        model="runware:108@22",  # Qwen model with fbCache support
        numberResults=1,
        height=1024,
        width=1024,
        acceleratorOptions=IAcceleratorOptions(
            fbcache=True,           # Enable First Block cache
            cacheStartStep=0,       # Start caching from step 0
            cacheStopStep=8         # Stop caching at step 8
        ),
    )

    images = await runware.imageInference(requestImage=request_image)
    for image in images:
        print(f"Image URL: {image.imageURL}")

teaCache

• teaCache is a boolean that enables or disables the teaCache feature. If set to True, it will use teaCache for faster inference.
  • It is specific to transformer models, Flux and SD3. teaCache does not work for UNet models like SDXL or SD1x.
• teaCacheDistance is a float between 0.0 and 1.0, where 0.0 is the most conservative and 1.0 is the most aggressive.
• cacheStartStep and cacheStopStep are integers that represent the start and end steps of the teaCache and DeepCache process.
  • cacheStartStep is the step at which the generator starts to skip blocks and reduce quality; cacheStopStep is the step at which the teaCache process ends, returning to full fidelity steps.
  • If not specified, teaCache (or DeepCache) will be enabled throughout the entire image generation process, which may be undesirable for preserving quality.

deepCache

• deepCache is a boolean that enables or disables the deepCache feature. If set to True, it will use deepCache for faster inference.
• deepCacheInterval represents the frequency of feature caching, specified as the number of steps between each cache operation.
  • A larger cache interval makes inference faster, and costs more quality.
  • The default value is 3
• deepCacheBranchId represents which branch of the network (ordered from the shallowest to the deepest layer) is responsible for executing the caching processes.
  • Opting for a lower branch ID will result in a more aggressive caching process, while a higher branch ID will yield a more conservative approach.
  • The default value is 0

fbCache

• fbcache is a boolean that enables or disables the First Block cache feature. If set to True, it will use fbCache for faster inference.
  • fbCache is compatible with specific models and provides additional acceleration options.
  • Works in conjunction with cacheStartStep and cacheStopStep to control the caching behavior.
• cacheStartStep and cacheStopStep control the range of steps where caching is applied.
  • cacheStartStep: The step at which caching begins (default: 0)
  • cacheStopStep: The step at which caching ends (default: total steps)
  • These parameters allow fine-grained control over when caching is active during the generation process.

Asynchronous Processing with Webhooks

The Runware SDK supports asynchronous processing via webhooks for long-running operations. When you provide a webhookURL, the API immediately returns a task response and sends the final result to your webhook endpoint when processing completes.

How it works

1. Include webhookURL parameter in your request
2. Receive immediate response with taskType and taskUUID
3. Final result is POSTed to your webhook URL when ready

Supported operations:

• Image Inference
• Photo Maker
• Image Caption
• Image Background Removal
• Image Upscale
• Prompt Enhance
• Video Inference

Example

from runware import Runware, IImageInference

async def main() -> None:
    runware = Runware(api_key=RUNWARE_API_KEY)
    await runware.connect()

    request_image = IImageInference(
        positivePrompt="a beautiful mountain landscape",
        model="civitai:36520@76907",
        height=512,
        width=512,
        webhookURL="https://your-server.com/webhook/runware"
    )

    # Returns immediately with task info
    response = await runware.imageInference(requestImage=request_image)
    print(f"Task Type: {response.taskType}")
    print(f"Task UUID: {response.taskUUID}")
    # Result will be sent to your webhook URL

Webhook Payload Format

Your webhook endpoint will receive a POST request with the same format as synchronous responses:

  "data": [
    {
      "taskType": "imageInference",
      "taskUUID": "a770f077-f413-47de-9dac-be0b26a35da6",
      "imageUUID": "77da2d99-a6d3-44d9-b8c0-ae9fb06b6200",
      "imageURL": "https://im.runware.ai/image/...",
      "cost": 0.0013
    }
  ]
}

Video Inference with Skip Response

For long-running video generation tasks, you can use skipResponse to submit the task and retrieve results later. This is useful for handling system interruptions, batch processing, or building queue-based systems.

from runware import Runware, IVideoInference

async def main() -> None:
    runware = Runware(api_key=RUNWARE_API_KEY)
    await runware.connect()

    # Submit video task without waiting
    request = IVideoInference(
            model="openai:3@2",
            positivePrompt="A beautiful sunset over the ocean",
            duration=4,
            width=1280,
            height=720,
            skipResponse=True,
    )

    response = await runware.videoInference(requestVideo=request)
    task_uuid = response.taskUUID
    print(f"Task submitted: {task_uuid}")
    
    # Later, retrieve results
    videos = await runware.getResponse(
        taskUUID=task_uuid,
        numberResults=1
    )
    
    for video in videos:
        print(f"Video URL: {video.videoURL}")

Parameters:

• skipResponse: Set to True to return immediately with taskUUID instead of waiting for completion
• Use getResponse(taskUUID) to retrieve results at any time

Video Inference with Async Delivery Method

For long-running video generation tasks, you can use deliveryMethod="async" to submit the task and retrieve results later. This is useful for handling system interruptions, batch processing, or building queue-based systems.

from runware import Runware, IVideoInference

async def main() -> None:
    runware = Runware(api_key=RUNWARE_API_KEY)
    await runware.connect()

    # Submit video task with async delivery method
    request = IVideoInference(
            model="openai:3@2",
            positivePrompt="A beautiful sunset over the ocean",
            duration=4,
            width=1280,
            height=720,
            deliveryMethod="async",
    )

    response = await runware.videoInference(requestVideo=request)
    task_uuid = response.taskUUID
    print(f"Task submitted: {task_uuid}")
    
    # Later, retrieve results
    videos = await runware.getResponse(
        taskUUID=task_uuid,
        numberResults=1
    )
    
    for video in videos:
        print(f"Video URL: {video.videoURL}")

Parameters:

• deliveryMethod: Set to "async" to return immediately with IAsyncTaskResponse containing taskUUID instead of waiting for completion
• Use getResponse(taskUUID) to retrieve results at any time
• deliveryMethod="sync" waits for complete results (may timeout for long-running tasks)

Enhancing Prompts

To enhance prompts using the Runware API, you can use the promptEnhance method of the Runware class. Here's an example:

from runware import Runware, IPromptEnhance

async def main() -> None:
    runware = Runware(api_key=RUNWARE_API_KEY)
    await runware.connect()

    prompt = "A beautiful sunset over the mountains"
    prompt_enhancer = IPromptEnhance(
        prompt=prompt,
        promptVersions=3,
        promptMaxLength=64,
    )

    enhanced_prompts = await runware.promptEnhance(promptEnhancer=prompt_enhancer)
    for enhanced_prompt in enhanced_prompts:
        print(enhanced_prompt.text)

Removing Image Background

To remove the background from an image using the Runware API, you can use the imageBackgroundRemoval method of the Runware class. Here's an example:

from runware import Runware, IImageBackgroundRemoval

async def main() -> None:
    runware = Runware(api_key=RUNWARE_API_KEY)
    await runware.connect()

    image_path = "image.jpg"
    remove_image_background_payload = IImageBackgroundRemoval(inputImage=image_path)

    processed_images = await runware.imageBackgroundRemoval(
        removeImageBackgroundPayload=remove_image_background_payload
    )
    for image in processed_images:
        print(image.imageURL)

Image-to-Text Conversion

To convert an image to text using the Runware API, you can use the imageCaption method of the Runware class. Here's an example:

from runware import Runware, IImageCaption

async def main() -> None:
    runware = Runware(api_key=RUNWARE_API_KEY)
    await runware.connect()

    image_path = "image.jpg"
    request_image_to_text_payload = IImageCaption(inputImage=image_path)

    image_to_text = await runware.imageCaption(
        requestImageToText=request_image_to_text_payload
    )
    print(image_to_text.text)

Video Caption

To generate captions for videos using the Runware API, you can use the videoCaption method of the Runware class. The SDK automatically polls for results when using async delivery. Here's an example:

from runware import Runware, IVideoCaption, IVideoCaptionInputs

async def main() -> None:
    runware = Runware(api_key=RUNWARE_API_KEY)
    await runware.connect()

    request_caption = IVideoCaption(
        model="memories:1@1",
        inputs=IVideoCaptionInputs(
            video="https://example.com/video.mp4"
        ),
        deliveryMethod="async",
        includeCost=True
    )

    caption_response = await runware.videoCaption(
        requestVideoCaption=request_caption
    )
    print(f"Caption: {caption_response.text}")
    if caption_response.cost:
        print(f"Cost: {caption_response.cost}")

Video Caption Parameters:

• model: Caption model identifier (e.g., "memories:1@1")
• inputs: IVideoCaptionInputs containing the video URL or UUID
• deliveryMethod: "async" (with automatic polling) or use webhookURL for webhook delivery
• includeCost: Include cost information in the response (optional)
• webhookURL: Webhook URL for async delivery without polling (optional)

Video Background Removal

To remove the background from videos you can use the videoBackgroundRemoval method of the Runware class. The SDK automatically polls for results when using async delivery. Here's an example:

from runware import Runware, IVideoBackgroundRemoval, IVideoBackgroundRemovalInputs, IVideoBackgroundRemovalSettings

async def main() -> None:
    runware = Runware(api_key=RUNWARE_API_KEY)
    await runware.connect()

    request_bg_removal = IVideoBackgroundRemoval(
        model="bria:51@1",
        inputs=IVideoBackgroundRemovalInputs(
            video="https://example.com/video.mp4"
        ),
        outputFormat="WEBM",
        includeCost=True,
        settings=IVideoBackgroundRemovalSettings(
            rgba=[255, 255, 255, 0]  
        )
    )

    processed_videos = await runware.videoBackgroundRemoval(
        requestVideoBackgroundRemoval=request_bg_removal
    )
    for video in processed_videos:
        print(f"Video URL: {video.videoURL}")
        if video.cost:
            print(f"Cost: {video.cost}")

Video Background Removal Parameters:

• model: Background removal model identifier (e.g., "bria:51@1")
• inputs: IVideoBackgroundRemovalInputs containing the video URL or UUID
• outputFormat: Output video format ("WEBM", "MP4", etc.)
• includeCost: Include cost information in the response (optional)
• settings: IVideoBackgroundRemovalSettings for custom background configuration
• webhookURL: Webhook URL for async delivery without polling (optional)

Background Removal Settings:

• rgba: Background color as [R, G, B, A] array (0-255 for RGB, 0.0-1.0 for alpha)

Upscaling Images

To upscale an image using the Runware API, you can use the imageUpscale method of the Runware class. Here's an example:

from runware import Runware, IImageUpscale

async def main() -> None:
    runware = Runware(api_key=RUNWARE_API_KEY)
    await runware.connect()

    image_path = "image.jpg"
    upscale_factor = 4

    upscale_gan_payload = IImageUpscale(
        inputImage=image_path, upscaleFactor=upscale_factor
    )
    upscaled_images = await runware.imageUpscale(upscaleGanPayload=upscale_gan_payload)
    for image in upscaled_images:
        print(image.imageURL)

Photo Maker

Use the photoMaker method of the Runware class. Here's an example:

from runware import Runware, IPhotoMaker
import uuid

async def main() -> None:
    runware = Runware(api_key=RUNWARE_API_KEY)
    await runware.connect()

    request_image = IPhotoMaker(
        model="civitai:139562@344487",
        positivePrompt="img of a beautiful lady in a forest",
        steps=35,
        numberResults=1,
        height=512,
        width=512,
        style="No style",
        strength=40,
        outputFormat="WEBP",
        includeCost=True,
        taskUUID=str(uuid.uuid4()),
        inputImages=[
            "https://im.runware.ai/image/ws/0.5/ii/74723926-22f6-417c-befb-f2058fc88c13.webp",
            "https://im.runware.ai/image/ws/0.5/ii/64acee31-100d-4aa1-a47e-6f8b432e7188.webp",
            "https://im.runware.ai/image/ws/0.5/ii/1b39b0e0-6bf7-4c9a-8134-c0251b5ede01.webp",
            "https://im.runware.ai/image/ws/0.5/ii/f4b4cec3-66d9-4c02-97c5-506b8813182a.webp"
        ],
    )
    
    
     photos = await runware.photoMaker(requestPhotoMaker=request_image)
     for photo in photos:
         print(f"Image URL: {photo.imageURL}")

ACE++

ACE++ (Advanced Character Edit) is an advanced framework for character-consistent image generation and editing. It allows you to create new images from a single reference image while preserving identity, and edit existing images without retraining the model.

Note: When using ACE++, you must set the model parameter to runware:102@1.

Character-Consistent Generation

To generate new images while preserving character identity from a reference image:

from runware import Runware, IImageInference, IAcePlusPlus

async def main() -> None:
    runware = Runware(api_key=RUNWARE_API_KEY)
    await runware.connect()

    # Upload your reference image first
    reference_image = await runware.uploadImage("path/to/reference_image.jpg")

    request_image = IImageInference(
        positivePrompt="photo of man wearing a business suit in a modern office",
        model="runware:102@1",                        # Required model for ACE++
        height=1024,
        width=1024,
        numberResults=1,
        acePlusPlus=IAcePlusPlus(
            inputImages=[reference_image.imageUUID],  # Reference image for character identity
            repaintingScale=0.3                       # Lower values (0.0-0.5) preserve more identity
        )
    )

    images = await runware.imageInference(requestImage=request_image)
    for image in images:
        print(f"Image URL: {image.imageURL}")

Character-Consistent Editing

To edit existing images while preserving character identity using masks:

from runware import Runware, IImageInference, IAcePlusPlus

async def main() -> None:
    runware = Runware(api_key=RUNWARE_API_KEY)
    await runware.connect()

    # Upload your reference image and mask
    reference_image = await runware.uploadImage("path/to/reference_image.jpg")
    mask_image = await runware.uploadImage("path/to/mask_image.png")

    request_image = IImageInference(
        positivePrompt="photo of woman wearing a red dress",
        model="runware:102@1",  # Required model for ACE++
        height=1024,
        width=1024,
        numberResults=1,
        acePlusPlus=IAcePlusPlus(
            inputImages=[reference_image.imageUUID],  # Reference image
            inputMasks=[mask_image.imageUUID],  # Mask for selective editing
            repaintingScale=0.7  # Higher values (0.5-1.0) follow prompt more in edited areas
        )
    )

    images = await runware.imageInference(requestImage=request_image)
    for image in images:
        print(f"Image URL: {image.imageURL}")

ACE++ Parameters:

• inputImages: Array containing exactly one reference image (required)
• inputMasks: Array containing at most one mask image (optional, for editing)
• repaintingScale: Float between 0.0 and 1.0
  • 0.0: Maximum character identity preservation
  • 1.0: Maximum adherence to prompt instructions
  • For generation: Use 0.0-0.5 for strong resemblance
  • For editing: Use 0.5-1.0 for more creative freedom in edited areas

Mask Requirements:

• The mask should be a black and white image
• White (255) represents areas to be edited
• Black (0) represents areas to be preserved
• Supported formats: PNG, JPG, WEBP

Generating Images with refiner

To generate images using the Runware API with refiner support, you can use the imageInference method of the Runware class. Here's an example:

from runware import Runware, IImageInference, IRefiner

async def main() -> None:
    runware = Runware(api_key=RUNWARE_API_KEY)
    await runware.connect()
    
    refiner = IRefiner(
        model="civitai:101055@128080",
        startStep=2,
        startStepPercentage=None,
    )

    request_image = IImageInference(
        positivePrompt="a beautiful sunset over the mountains",
        model="civitai:101055@128078",
        numberResults=4,
        negativePrompt="cloudy, rainy",
        height=512,
        width=512,
        refiner=refiner
    )

    images = await runware.imageInference(requestImage=request_image)
    for image in images:
        print(f"Image URL: {image.imageURL}")

Using ControlNet with Image Inference

To use ControlNet for image inference in the Runware SDK, you can use a class IControlNetGeneral. Here's an example of how to set up and use this feature:

from runware import Runware, IImageInference, IControlNetGeneral, EControlMode

async def main() -> None:
    runware = Runware(api_key=RUNWARE_API_KEY)
    await runware.connect()

    controlNet = IControlNetGeneral(
        startStep=1,
        endStep=30,
        weight=0.5,
        controlMode=EControlMode.BALANCED.value,
        guideImage="https://huggingface.co/datasets/mishig/sample_images/resolve/main/canny-edge.jpg",
        model='civitai:38784@44716'
    )

    request_image = IImageInference(
        positivePrompt="a beautiful sunset",
        model='civitai:4384@128713',
        controlNet=[controlNet],
        numberResults=1,
        height=512,
        width=512,
        outputType="URL",
        seed=1568,
        steps=40
    )

    images = await runware.imageInference(requestImage=request_image)

    for image in images:
        print(f"Image URL: {image.imageURL}")

This example demonstrates how to configure and use a ControlNet to enhance the image inference process.

Inferencing Ace++ Pipeline

To use Ace++ in the Runware SDK, you can use a class IAcePlusPlus. Here's an example of how to set up and use this feature: Much more examples are in examples/ace++

from runware import Runware, IImageInference, IAcePlusPlus

async def main() -> None:
    runware = Runware(api_key=RUNWARE_API_KEY)
    await runware.connect()

    # Upload your reference image and mask
    reference_image = "https://raw.githubusercontent.com/ali-vilab/ACE_plus/refs/heads/main/assets/samples/application/logo_paste/1_ref.png"
    mask_image = "https://raw.githubusercontent.com/ali-vilab/ACE_plus/refs/heads/main/assets/samples/application/logo_paste/1_1_m.png"
    init_image = "https://raw.githubusercontent.com/ali-vilab/ACE_plus/refs/heads/main/assets/samples/application/logo_paste/1_1_edit.png"
    request_image = IImageInference(
        positivePrompt="The logo is printed on the headphones.",
        model="runware:102@1",  # Required model for ACE++
        taskUUID="68020b8f-bbcf-4779-ba51-4f3bb00aef6a",
        height=1024,
        width=1024,
        numberResults=1,
        steps=28,
        CFGScale=50.0,
        referenceImages=[reference_image],  # Reference image
        acePlusPlus=IAcePlusPlus(
            inputImages=[init_image],  # Input image
            inputMasks=[mask_image],  # Mask for selective editing
            repaintingScale=1.0,
            taskType="subject"  # Can be one of "portrait", "subject", "local_editing"
        ),
    )
    print(f"Sending request: {request_image}")
    images = await runware.imageInference(requestImage=request_image)
    
    for image in images:
        print(f"Image URL: {image.imageURL}")

This example demonstrates how to configure and use a ControlNet to enhance the image inference process.

Generating Images with OpenAI Models (DALL-E 2 & DALL-E 3)

The Runware SDK supports OpenAI's DALL-E 2 and DALL-E 3 models for image generation. These models offer high-quality image generation with various configuration options.

DALL-E 2

from runware import Runware, IImageInference, IOpenAIProviderSettings

async def main() -> None:
    runware = Runware(api_key=RUNWARE_API_KEY)
    await runware.connect()

    # DALL-E 2 configuration
    provider_settings = IOpenAIProviderSettings(
        quality="high",
        background="transparent"  # Optional: for transparent backgrounds
    )

    request_image = IImageInference(
        positivePrompt="A cute cartoon robot character",
        model="openai:1@1",  # DALL-E 2 model identifier
        width=1024,
        height=1024,
        numberResults=1,
        outputFormat="PNG",
        includeCost=True,
        providerSettings=provider_settings
    )

    images = await runware.imageInference(requestImage=request_image)
    for image in images:
        print(f"Image URL: {image.imageURL}")

DALL-E 3

from runware import Runware, IImageInference, IOpenAIProviderSettings

async def main() -> None:
    runware = Runware(api_key=RUNWARE_API_KEY)
    await runware.connect()

    # DALL-E 3 with HD quality
    provider_settings = IOpenAIProviderSettings(
        quality="hd"  # Options: "hd" or "standard"
    )

    request_image = IImageInference(
        positivePrompt="A futuristic city with flying cars, highly detailed",
        model="openai:2@3",  # DALL-E 3 model identifier
        width=1024,
        height=1024,
        numberResults=1,
        outputFormat="PNG",
        includeCost=True,
        providerSettings=provider_settings
    )

    images = await runware.imageInference(requestImage=request_image)
    for image in images:
        print(f"Image URL: {image.imageURL}")

OpenAI Provider Settings:

• quality: Image quality setting
  • DALL-E 2: "high" (recommended)
  • DALL-E 3: "hd" or "standard"
• background: (DALL-E 2 only) Set to "transparent" for transparent backgrounds
• style: (Optional) Additional style parameters

Model Identifiers:

• DALL-E 2: "openai:1@1"
• DALL-E 3: "openai:2@3"

Inferencing Video Models

To inference Video Generation Models in the Runware SDK, you can use a class IVideoInference. Almost every video model support its own providerSettings: IMinimaxProviderSettings, IBytedanceProviderSettings, IGoogleProviderSettings, IKlingAIProviderSettings, IPixverseProviderSettings, IViduProviderSettings. More examples can be found in examples/video.

Here's an example of an image-to-video (i2v) task using Google's Veo3:

import asyncio
from runware import Runware, IVideoInference, IGoogleProviderSettings, IFrameImage


async def main():
    runware = Runware(
        api_key=RUNWARE_API_KEY,
    )
    await runware.connect()

    request = IVideoInference(
        positivePrompt="spinning galaxy",
        model="google:3@0",
        width=1280,
        height=720,
        numberResults=1,
        seed=10,
        includeCost=True,
        frameImages=[ # Comment this to use t2v
            IFrameImage(
                inputImage="https://github.com/adilentiq/test-images/blob/main/common/image_15_mb.jpg?raw=true",
            ),
        ],
        providerSettings=IGoogleProviderSettings(
            generateAudio=True,
            enhancePrompt=True
        )
    )
    videos = await runware.videoInference(requestVideo=request)
    for video in videos:
        print(f"Video URL: {video.videoURL}")
        print(f"Cost: {video.cost}")
        print(f"Seed: {video.seed}")
        print(f"Status: {video.status}")


if __name__ == "__main__":
    asyncio.run(main())

Audio Inference

To generate audio using the Runware SDK, you can use the audioInference method with the IAudioInference class. The SDK supports various audio generation models including ElevenLabs and other providers.

Here's an example of generating audio using ElevenLabs:

import asyncio
from runware import Runware, IAudioInference, IAudioSettings

async def main() -> None:
    runware = Runware(api_key=RUNWARE_API_KEY)
    await runware.connect()

    # Create audio settings
    audio_settings = IAudioSettings(
        sampleRate=22050,  # Sample rate in Hz
        bitrate=32         # Audio bitrate
    )

    # Create audio inference request
    request_audio = IAudioInference(
        model="elevenlabs:1@1",           # ElevenLabs model
        positivePrompt="upbeat electronic music with synthesizers and drums",
        outputFormat="MP3",               # Output format: MP3, WAV, etc.
        outputType="URL",                 # Return URL or base64
        audioSettings=audio_settings,
        numberResults=1,                  # Number of audio files to generate
        duration=10,                      # Duration in seconds
        includeCost=True                  # Include cost information
    )

    
    audio_results = await runware.audioInference(requestAudio=request_audio)
    for audio in audio_results:
        print(f"Audio URL: {audio.audioURL}")
        print(f"Duration: {audio.duration}")
        print(f"Cost: {audio.cost}")


if __name__ == "__main__":
    asyncio.run(main())

Audio Settings

The IAudioSettings class allows you to configure audio generation parameters:

• sampleRate: Audio sample rate in Hz (e.g., 22050, 44100)
• bitrate: Audio bitrate for compressed formats

Audio Inference Parameters

The IAudioInference class supports the following parameters:

• model: Audio generation model identifier (e.g., "elevenlabs:1@1")
• positivePrompt: Text description of the audio to generate
• outputFormat: Output audio format ("MP3", "WAV", etc.)
• outputType: Return type ("URL" or "BASE64")
• audioSettings: Audio configuration settings
• numberResults: Number of audio files to generate
• duration: Duration of the generated audio in seconds
• includeCost: Whether to include cost information in the response

Model Upload

To upload model using the Runware API, you can use the uploadModel method of the Runware class. Here are examples:

from runware import Runware, IUploadModelCheckPoint


async def main() -> None:
    runware = Runware(api_key=RUNWARE_API_KEY)
    await runware.connect()

    payload = IUploadModelCheckPoint(
        air='qatests:68487@08629',
        name='yWO8IaKwez',
        heroImageURL='https://raw.githubusercontent.com/adilentiq/test-images/refs/heads/main/image.jpg',
        downloadURL='https://repo-controlnets-r2.runware.ai/controlnet-zoe-depth-sdxl-1.0.safetensors'
                    '/controlnet-zoe-depth-sdxl-1.0.safetensors.part-001-1',
        uniqueIdentifier='aq2w3e4r5t6y7u8i9o0p1q2w3e4r5t6y7u8i9o0p1q2w3e4r5t6y7u8i9o0p1234',
        version='1.0',
        tags=['tag1', 'tag2', 'tag2'],
        architecture='flux1d',
        type='base',
        defaultWeight=0.8,
        format='safetensors',
        positiveTriggerWords='my trigger word',
        shortDescription='a model description',
        private=False,
        defaultScheduler='Default',
        comment='some comments if you want to add for internal use',
    )

    uploaded = await runware.modelUpload(payload)
    print(f"Response : {uploaded}")

from runware import Runware, IUploadModelLora


async def main() -> None:
    runware = Runware(api_key=RUNWARE_API_KEY)
    await runware.connect()

    payload = IUploadModelLora(
        air='qatests:68487@08629',
        name='yWO8IaKwez',
        heroImageURL='https://raw.githubusercontent.com/adilentiq/test-images/refs/heads/main/image.jpg',
        downloadURL='https://repo-controlnets-r2.runware.ai/controlnet-zoe-depth-sdxl-1.0.safetensors'
                    '/controlnet-zoe-depth-sdxl-1.0.safetensors.part-001-1',
        uniqueIdentifier='aq2w3e4r5t6y7u8i9o0p1q2w3e4r5t6y7u8i9o0p1q2w3e4r5t6y7u8i9o0p1234',
        version='1.0',
        tags=['tag1', 'tag2', 'tag2'],
        architecture='flux1d',
        defaultWeight=0.8,
        format='safetensors',
        positiveTriggerWords='my trigger word',
        shortDescription='a model description',
        private=False,
        comment='some comments if you want to add for internal use',
    )

    uploaded = await runware.modelUpload(payload)
    print(f"Response : {uploaded}")

from runware import Runware, IUploadModelControlNet


async def main() -> None:
    runware = Runware(api_key=RUNWARE_API_KEY)
    await runware.connect()

    payload = IUploadModelControlNet(
        air='qatests:68487@08629',
        name='yWO8IaKwez',
        heroImageURL='https://raw.githubusercontent.com/adilentiq/test-images/refs/heads/main/image.jpg',
        downloadURL='https://repo-controlnets-r2.runware.ai/controlnet-zoe-depth-sdxl-1.0.safetensors'
                    '/controlnet-zoe-depth-sdxl-1.0.safetensors.part-001-1',
        uniqueIdentifier='aq2w3e4r5t6y7u8i9o0p1q2w3e4r5t6y7u8i9o0p1q2w3e4r5t6y7u8i9o0p1234',
        version='1.0',
        tags=['tag1', 'tag2', 'tag2'],
        architecture='flux1d',
        format='safetensors',
        shortDescription='a model description',
        private=False,
        comment='some comments if you want to add for internal use',
    )


uploaded = await runware.modelUpload(payload)
print(f"Response : {uploaded}")

Image Background Removal

There are two ways to remove the background from an image.

1. Using the settings parameter of the IImageBackgroundRemoval class.
2. Without using the settings parameter and using the model parameter to specify the model to use.

Using the settings parameter

Note: When using the rgba parameter, the final a value is a float between 0.0 and 1.0, but a value of 1-255 will be internally scaled down to the correct float range.

from runware import Runware, RunwareAPIError, IImage, IImageBackgroundRemoval, IBackgroundRemovalSettings
import asyncio
import os
from dotenv import load_dotenv

load_dotenv(override=True)


async def main() -> None:
    runware = Runware(
        api_key=os.environ.get("RUNWARE_API_KEY")
    )
    await runware.connect()
    background_removal_settings = IBackgroundRemovalSettings(
        rgba=[255, 255, 255, 0],
        alphaMatting=True,
        postProcessMask=True,
        returnOnlyMask=False,
        alphaMattingErodeSize=10,
        alphaMattingForegroundThreshold=240,
        alphaMattingBackgroundThreshold=10
        )

    request_image = IImageBackgroundRemoval(
        taskUUID="abcdbb9c-3bd3-4d75-9234-bffeef994772",
        inputImage="https://raw.githubusercontent.com/adilentiq/test-images/refs/heads/main/common/headphones.jpeg",
        settings=background_removal_settings,
        outputType="URL",
        outputFormat="PNG",
        includeCost=True,
    )

    print(f"Payload: {request_image}")
    try:
        processed_images: List[IImage] = await runware.imageBackgroundRemoval(
            removeImageBackgroundPayload=request_image
        )
    except RunwareAPIError as e:
        print(f"API Error: {e}")
        print(f"Error Code: {e.code}")
    except Exception as e:
        print(f"Unexpected Error: {e}")
    else:
        print("Processed Image with the background removed:")
        print(processed_images)
        for image in processed_images:
            print(image.imageURL)


asyncio.run(main())

Using the model parameter

from runware import Runware, RunwareAPIError, IImage, IImageBackgroundRemoval
import asyncio
import os
from dotenv import load_dotenv

load_dotenv(override=True)


async def main() -> None:
    runware = Runware(
        api_key=os.environ.get("RUNWARE_API_KEY"),
    )
    await runware.connect()

    request_image = IImageBackgroundRemoval(
        taskUUID="abcdbb9c-3bd3-4d75-9234-bffeef994772",
        model="runware:110@1",
        inputImage="https://raw.githubusercontent.com/adilentiq/test-images/refs/heads/main/common/headphones.jpeg"
    )

    print(f"Payload: {request_image}")
    try:
        processed_images: List[IImage] = await runware.imageBackgroundRemoval(
            removeImageBackgroundPayload=request_image
        )
    except RunwareAPIError as e:
        print(f"API Error: {e}")
        print(f"Error Code: {e.code}")
    except Exception as e:
        print(f"Unexpected Error: {e}")
    else:
        print("Processed Image with the background removed:")
        print(processed_images)
        for image in processed_images:
            print(image.imageURL)


asyncio.run(main())

For more detailed usage and additional examples, please refer to the examples directory.

Configuring Timeouts

The Runware SDK provides configurable timeout settings for different operations through environment variables. All timeout values are in milliseconds.

Timeout Configuration

Set environment variables to customize timeout behavior:

# Concurrency
RUNWARE_MAX_CONCURRENT_REQUESTS=15           # Max concurrent API requests per client (default: 15)

# Image Operations (milliseconds)
RUNWARE_IMAGE_INFERENCE_TIMEOUT=300000      # Image generation (default: 5 min)
RUNWARE_IMAGE_OPERATION_TIMEOUT=120000      # Caption, upscale, background removal (default: 2 min)
RUNWARE_IMAGE_UPLOAD_TIMEOUT=60000          # Image upload (default: 1 min)
RUNWARE_MAX_POLLS_IMAGE_GENERATION=480      # Max polling attempts for async image / getResponse (default: 480)

# Model Operations (milliseconds)
RUNWARE_MODEL_UPLOAD_TIMEOUT=900000         # Model upload (default: 15 min)

# Video Operations (milliseconds)
RUNWARE_VIDEO_INITIAL_TIMEOUT=30000         # Initial response wait (default: 30 sec)
RUNWARE_VIDEO_POLLING_DELAY=3000            # Delay between status checks (default: 3 sec)
RUNWARE_MAX_POLLS_VIDEO_GENERATION=480      # Max polling attempts for video / caption / upscale / bg removal (default: 480, ~24 min total)

# 3D Operations
RUNWARE_MAX_POLLS_3D_GENERATION=480        # Max polling attempts for 3D inference / getResponse (default: 480)

# Audio Operations (milliseconds)
RUNWARE_AUDIO_INFERENCE_TIMEOUT=300000      # Audio generation (default: 5 min)
RUNWARE_AUDIO_POLLING_DELAY=1000            # Delay between status checks (default: 1 sec)
RUNWARE_MAX_POLLS_AUDIO_GENERATION=240      # Max polling attempts for audio inference (default: 240, ~4 min total)

# Other Operations (milliseconds)
RUNWARE_PROMPT_ENHANCE_TIMEOUT=60000        # Prompt enhancement (default: 1 min)
RUNWARE_WEBHOOK_TIMEOUT=30000               # Webhook acknowledgment (default: 30 sec)
RUNWARE_TIMEOUT_DURATION=480000             # General operations (default: 8 min)

Usage Example

import os

# Configure before importing Runware
os.environ["RUNWARE_VIDEO_POLLING_DELAY"] = "5000"  # 5 seconds between checks
os.environ["RUNWARE_MAX_POLLS_VIDEO_GENERATION"] = "600"  # Allow up to 50 minutes

from runware import Runware

async def main():
    runware = Runware(api_key=os.getenv("RUNWARE_API_KEY"))
    await runware.connect()
    # Your code here

Note: For long-running video operations, consider using webhooks or skipResponse=True to avoid timeout issues with extended generation times.