Skip to main content

​
What is Daydream Scope?

Daydream Scope is an open-source tool for running and customizing real-time interactive generative AI pipelines and models. It’s currently in alpha, so expect some rough edges, but we’re excited to iterate in public with the open-source AI community. Scope enables you to:
  • Stream real-time AI-generated video via WebRTC with low latency
  • Use an interactive timeline editor to modify generation parameters on the fly
  • Work with multi-modal inputs including text prompts, videos, camera feeds, and more
  • Experiment with state-of-the-art video diffusion models

​
Supported Pipelines

Scope currently supports three autoregressive video diffusion models:

StreamDiffusion

Real-time video generation with streaming capabilities for immediate visual feedback

LongLive

Extended generation capabilities for longer video sequences

Krea Realtime

Text-to-video generation with real-time streaming (requires ≥32GB VRAM)
Krea Realtime requirements: NVIDIA GPU with ≥32GB VRAM (≥40GB for higher resolutions). Can run on 32GB with fp8 quantization at lower resolutions.

​
Choose Your Deployment Path

Pick the option that best fits your setup:

Local Installation

Best performance with NVIDIA RTX 4090/5090 or similar GPU

RunPod Cloud

No GPU required - deploy in minutes using our template

​
Local Installation

Best for: Users with high-end NVIDIA GPUs who want maximum performance and full control

​
System Requirements

  • OS: Linux or Windows
  • GPU: NVIDIA RTX 4090/5090 or similar with ≥24GB VRAM
  • Drivers: CUDA 12.8+
Newer GPU generations provide higher FPS throughput and lower latency. Make sure your GPU has at least 24GB of VRAM and supports CUDA 12.8 or higher.
Krea Realtime requirements: If you plan to use the Krea Realtime pipeline, you’ll need ≥32GB VRAM (≥40GB for higher resolutions). You can run it on 32GB with fp8 quantization at lower resolutions.

​
Installation Steps

1

Check your GPU drivers

First, verify that your NVIDIA drivers are properly installed and support CUDA 12.8 or higher:
nvidia-smi
The output should show your GPU and a CUDA version of at least 12.8.
2

Install dependencies

You’ll need the following installed on your system:
  • UV (Python package manager used to run the server)
  • Node.js and npm
If you don’t have these installed, visit their respective websites for installation instructions.
3

Clone the repository

Clone the Scope repository to your local machine:
git clone git@github.com:daydreamlive/scope.git
cd scope
4

Build frontend and install Python dependencies

Run the build command to set up both the frontend and backend:
uv run build
This will install all required Python packages including Torch and FlashAttention. The first-time install may take a while as dependencies are downloaded and compiled.
5

Start the Scope server

Launch the Scope server:
uv run daydream-scope
On the first run, model weights will download automatically to ~/.daydream-scope/models. This may take some time depending on your internet connection.
6

Access the UI

Once the server is running, open your browser and navigate to:http://localhost:8000You should see the Scope interface ready to use!

​
Cloud Deployment (RunPod)

Best for: Researchers and developers without access to local high-end GPUs
RunPod is a third-party cloud GPU service that lets you run Scope without any local hardware requirements. We’ve created a template to make deployment as simple as possible. You’ll need a RunPod account with credits added. Create an account at runpod.io and add funds to get started. Pricing varies based on GPU selection - check RunPod’s pricing page for current rates. Prefer video? Watch the tutorial below or follow the written steps.

​
Deployment Steps

1

Access the Daydream Scope template

Click the link below to access our pre-configured RunPod template:RunPod Template: Deploy Daydream Scope on RunPodThis will take you directly to the RunPod deployment console with our template loaded.
2

Create a HuggingFace token

RunPod deployment requires a HuggingFace token to enable TURN server functionality. This helps establish WebRTC connections in cloud environments with restrictive firewall settings.
  1. Create a free account at huggingface.co (if you don’t have one)
  2. Navigate to Settings → Access Tokens
  3. Click New token and create a token with read permissions
  4. Copy the token - you’ll need it in the next step
The HuggingFace integration provides 10GB of free streaming per month via Cloudflare TURN servers.
3

Select your GPU

Choose a GPU that meets Scope’s requirements:
  • Minimum: ≥24GB VRAM
  • Recommended: NVIDIA RTX 4090/5090 or similar
  • Drivers: CUDA 12.8+ support
Newer GPU generations will give you better FPS and lower latency for a smoother experience.
4

Configure environment variables

Now you’ll add your HuggingFace token to the deployment:
  1. Click “Edit Template” in the RunPod interface
  2. Find the environment variables section
  3. Add a new variable named HF_TOKEN
  4. Paste your HuggingFace token as the value
  5. Click “Save” to save your changes
This token enables the TURN server for reliable WebRTC streaming in cloud environments.
5

Deploy your instance

Click “Deploy On-Demand” to start your RunPod instance.Wait for the deployment to complete. This usually takes a few minutes as the container initializes and downloads necessary model weights.
6

Access your Scope instance

Once deployment is complete, RunPod will provide you with a URL. Open the app at port 8000.Your URL will look something like: https://your-instance-id.runpod.io:8000The Scope interface should now be ready to use!

​
Using Scope

Now that you have Scope running (either locally or on RunPod), here’s what to expect and how to get started.

​
Your First Run

When you first open Scope, you’ll see:
  • Default mode: Video mode with a looping cat test video
  • Default prompt: “a dog walking in grass”
  • Expected speed: ~8 FPS (varies depending on hardware)
Try updating the prompt in real-time! For example:
  • “a cow walking in grass”
  • “a dragon flying through clouds”
  • “a robot walking on mars”
You should see the video transform based on your prompt while maintaining the structure and motion of the original video.

​
Key Features

​
Video Mode

Apply your prompts to static test videos. This is perfect for experimenting and understanding how Scope transforms video content based on your text descriptions.

​
Camera Mode

Connect a camera to use live camera feeds as input. This enables real-time interactive experiences where you can see AI transformations happening live.

​
Interactive Timeline Editor

One of Scope’s most powerful features - modify generation parameters over time. You can:
  • Replay example generations
  • Modify prompts at different points in the timeline
  • Steer the generation in different directions
  • Import and export timeline files for reproducible workflows

​
Custom Prompts

Swap in different characters, scenes, styles, or entirely new concepts. Scope supports rich text-based controls for fine-tuning your generations.

​
Model Parameter Controls

Scope gives you access to various model parameters for fine-tuning generation behavior. Experiment with these to achieve different effects and optimize for your specific use case.

​
Troubleshooting

​
Local Installation Issues

  • Run nvidia-smi and verify CUDA version is ≥ 12.8
  • Update your NVIDIA drivers if needed
  • Ensure UV, Node.js, and npm are properly installed
  • Try clearing the cache and rebuilding: uv cache clean && uv run build
This error has been encountered on certain Linux machines when the Python header file is missing.Solution: Install the python3-dev package.On Debian/Ubuntu-based systems:
sudo apt-get install python3-dev
On other Linux distributions, install the equivalent Python development package for your system.
  • Check your internet connection
  • Verify you have sufficient disk space in ~/.daydreamscope/models
  • Model downloads can be large, so be patient on first run

​
RunPod Deployment Issues

  • Verify the instance is fully deployed (check RunPod dashboard)
  • Make sure you’re accessing port 8000
  • Check that your HF_TOKEN is correctly set in environment variables
  • Try selecting a more powerful GPU
  • Check your internet connection speed
  • The TURN server should help, but network conditions vary
  • Verify your HF_TOKEN is valid and has read permissions
  • Check that the token is properly set in the environment variables
  • Try redeploying the instance with the correct token

​
Next Steps

Now that you have Scope running, here’s how to dive deeper and connect with the community:

Join our Discord

Connect with the community, get help, and share your creations in our #scope channel

Contribute on GitHub

Scope is open-source - contribute code, report issues, or suggest features
Explore the different pipelines, experiment with the timeline editor, and see what creative applications you can build with real-time AI video generation!