Hey @nari-labs team! 👋

Dia is great! The way it generates dialogue with speaker tags and handles things like (laughs) is really well done.

To help people try Dia and use it in their projects, I've put it on Replicate:

Demo & API: https://replicate.com/zsxkib/dia

I also posted about it on X/Twitter here: https://x.com/zsakib_/status/1915037657064759716

What's new in this PR:

This PR adds the necessary files and configuration to run the existing Dia model on Replicate:

• Added predict.py: A script to load the model and run it using Cog's interface.
• Added cog.yaml: Defines the Python environment, system packages, and GPU requirements for Replicate.
• Added .dockerignore: Optimizes the Docker build by excluding unnecessary files.
• Updated .gitignore: Includes Replicate/Cog specific files like .cog/ and /model_cache/.
• Updated README.md: Added a Replicate badge linking to the demo page.

This setup provides:

1. A web interface for easy testing without any local setup.
2. An API endpoint for developers to run Dia generation from their own code.

Example API Usage:

import replicate

output = replicate.run(
    "zsxkib/dia:VERSION_HASH", # Replace VERSION_HASH with the actual model version hash
    input={
        "text": "[S1] Hello there! How are you? [S2] I'm doing great, thanks for asking! (laughs)",
        "audio_prompt": "https://replicate.delivery/pbxt/YourAudioFile.mp3", # Optional
        "max_audio_prompt_seconds": 10, # Optional, controls prompt length (default 10s)
        "temperature": 1.3,
        "cfg_scale": 3.0
        # ... other parameters ...
    }
)
print(output)
# Output: URL to the generated .wav file

On Replicate, users pay for their own compute time, which helps keep things running even if a model gets popular.

Let me know if you have any questions or feedback on the Replicate page. Happy to change anything.

Thanks for making Dia open source!

- Sakib @ Replicate