Spaces:

jbilcke-hf

/

AI-WebTV

like 487

Running

Thank you! I must admit I had way too much fun working on this 😁

what I like about it is that it combines multiple art forms, it will be a great showcase for future models

I second the prais and I'd love to read more about the setup. Seems like a non-trivial task to generate the videos and music in real time (unless it was all pre-computed and how is running on a loop).
Really enjoyable to watch, and with each demo like this I found that gradio can handle something I thought was not possible - didn't realise a video stream/full screen video is doable.

The new interpolation is very nice, much smoother!

Thank you @matthoffner

@anwo Here are some explanations about the architecture of the web tv.

The main code of the webtv is located inside the media-server :

manual steps:

• human input to write a short paragraph describing a multi-shot video sequence
• manual submit it to GPT-4 to generate a list of video captions for each shot (the system instructions are extracts from a stable diffusion guide)
• commit the captions to the playlist database

Inside the media-server space (generation process running in the background):

• for each prompt in the database
• generate a silent 3 seconds video clip with Zeroscope V2 576w (hosted on Hugging Face Spaces)
• upscale the clip with Zeroscope V2 XL (also a HF Space)
• perform frame interpolation with FILM (also a HF Space)
• storage in the Persistent Storage of the media-server Space

Inside the media-server space (streaming process running in the foreground):

• for each video file in the persistent storage folder
• add it to a new FFmpeg playlist (it's just a .txt file)
• broadcast it over the RTMP protocol using FFmpeg (in FLV format)
• diffusion of the stream using node-media-server

Inside the AI-WebTV space:

• display the stream using mpegts.js
• this doesn't work on iPhone, but now there is also a Twitch mirror