Sieve: Ship powerful AI features, fast.

Ship powerful AI features, fast.

Stop worrying about complex infrastructure and broken tools. Leverage the latest and greatest models, while collaborating with your entire team.

How Sieve WorksNew features in minutesThe AI space is moving so fast, it's hard to keep up. Siloed code, slow processing, and painful deployment should be a thing of the past. We built a cloud-native workflow orchestration tool that can be used by developers, PMs, and designers alike.

Deploy Workflow

Pick a starter template, leverage building blocks, deploy custom code, or use our visual editor.

Push Data

Trigger a workflow from the API, CLI, or automatically generated UI in Sieve's dashboard.

Iterate and Experiment

Test varying inputs, API providers, model versions, and more. Log results to track and improve performance.

Deploy to Production

Already production-ready. Integrate with a single API call, scale to thousands of jobs effortlessly.

API DemoStart with a templateStart using Sieve as a simple API using a variety of template workflows. Custom object detection, background blur and removal effects, text-based queries to search over videos, talking avatar generation, and more.

Step 1

Push a job

POST /v1/push

POST /v1/push

Step 2

Get the results

GET /v1/jobs/{id}

GET /v1/jobs/{id}

Send request to see results

Built for developers, designers, and PMs alikeBuild custom workflowsSieve is a fully-featured workflow orchestration tool that you can use to build for virtually any use case. Developers can use the SDK to deploy custom models, while designers and PMs can use the visual editor to quickly build custom workflows.

For engineers and data scientists

Step 1

Define models and functions

Just define some Python code, libraries, and dependencies.

@sieve.Model(
  name="u2netp",
  gpu = True,
  python_packages=["torch==1.8.1", "torchvision==0.9.1"],
  system_packages=["libgl1-mesa-glx", "libglib2.0-0", "ffmpeg"],
  python_version="3.8"
)
class U2Net:
  def __setup__(self):
      self.model = load_model()

  def __predict__(self, img: sieve.Image) -> sieve.Image:
      frame_data = cv2.cvtColor(img.array, cv2.COLOR_BGR2RGB)
      width = frame_data.shape[1]
      height = frame_data.shape[0]
      
      output_image = predict(self.model, frame_data)
      output_image = cv2.resize(output_image, (width, height), interpolation=cv2.INTER_CUBIC)
      return sieve.Image(array=output_image)

@sieve.Model(
  name="u2netp",
  gpu = True,
  python_packages=["torch==1.8.1", "torchvision==0.9.1"],
  system_packages=["libgl1-mesa-glx", "libglib2.0-0", "ffmpeg"],
  python_version="3.8"
)
class U2Net:
  def __setup__(self):
      self.model = load_model()

  def __predict__(self, img: sieve.Image) -> sieve.Image:
      frame_data = cv2.cvtColor(img.array, cv2.COLOR_BGR2RGB)
      width = frame_data.shape[1]
      height = frame_data.shape[0]
      
      output_image = predict(self.model, frame_data)
      output_image = cv2.resize(output_image, (width, height), interpolation=cv2.INTER_CUBIC)
      return sieve.Image(array=output_image)

Step 2

Define your workflow

@sieve.workflow(name="video_background_blur")
def background_blur(video: sieve.Video) -> Dict:
    images = VideoSplitter(video)
    masks = U2Net()(images)
    blurred_images = blur_background(images, masks)
    combiner = sieve.reference('sieve/frame-combiner')
    return combiner(blurred_images)

@sieve.workflow(name="video_background_blur")
def background_blur(video: sieve.Video) -> Dict:
    images = VideoSplitter(video)
    masks = U2Net()(images)
    blurred_images = blur_background(images, masks)
    combiner = sieve.reference('sieve/frame-combiner')
    return combiner(blurred_images)

Step 3

Deploy

We'll build and deploy your workflow to the cloud for you in minutes. No need to worry about infrastructure.

sieve deploy

sieve deploy

Step 4

Push Data

Push data to your workflow in a single API call.

POST https://mango.sievedata.com/v1/push

POST https://mango.sievedata.com/v1/push

For designers and PMs

Step 1

Browse template workflows and models

Pick a template workflow or build one using the visual editor.

Step 2

Drag and drop

Pick a template workflow or build one using the visual editor and the model library.

Step 3

Push Data

Auto-generated, typed UI to push data with no code.

Iterate and experiment with your entire teamExperiment and collaborate, fast.We built a framework that makes it easy to reference public or private models, either from code or the visual editor. Sieve is also updated regularly with the latest open-source models and workflows.

Example: AI Talking Head Avatar Generation

@sieve.workflow(name="talking_head_generator")
def talking_head(driving_video: sieve.Video, script: str, avatar: sieve.Image) -> sieve.Video:
    audio = sieve.reference('sieve/custom_tortoise_tts')(script)
    frames = sieve.reference('sieve/video-splitter')(driving_video)
    face_boxes = sieve.reference('sieve/face-detector')(frames)
    faces_tracked = sieve.reference('sieve/sort')(face_boxes)
    video = sieve.reference('sieve/wav2lip')(driving_video, audio, faces_tracked)
    talking_head = sieve.reference('sieve/thinplate-mm')(video, avatar)
    return talking_head

@sieve.workflow(name="talking_head_generator")
def talking_head(driving_video: sieve.Video, script: str, avatar: sieve.Image) -> sieve.Video:
    audio = sieve.reference('sieve/custom_tortoise_tts')(script)
    frames = sieve.reference('sieve/video-splitter')(driving_video)
    face_boxes = sieve.reference('sieve/face-detector')(frames)
    faces_tracked = sieve.reference('sieve/sort')(face_boxes)
    video = sieve.reference('sieve/wav2lip')(driving_video, audio, faces_tracked)
    talking_head = sieve.reference('sieve/thinplate-mm')(video, avatar)
    return talking_head

Built to scaleForget about infrastructure

Flexible inference

Start with serverless, always-available auto-scaling. Then, scale up to dedicated GPU instances for faster and more cost-effective inference.

Advanced optimizations

Even faster processing due to optimizations for rich data using specialized data storage and transfer via `sieve.Video`, `sieve.Image`, and `sieve.Audio` data types.

Stream and batch processing

Use Sieve's iterator support to stream results from jobs as they are processed, or launch massively parallel jobs to process large sets of data.

Best-in-class orchestration

Automatic retries, job monitoring dashboards, run logs, and everything else you'd expect from a workflow orchestration tool.

Build for your custom use caseUse Cases

Content Search

Object Detection

Highlight Generation

AI Avatars

Background Removal

Voice Cloning

Audio Enhancement

Rotoscoping

Art Generation

Audio Transcription

Scene Detection

Content Tagging

Content Upscaling

Style Transfer

Custom Workflow

Learn how Sieve can support your custom use case

Want to learn more?