Sign In

DiffusionTV: Watch Civitai Videos in Tiktok Format

DiffusionTV: Watch Civitai Videos in Tiktok Format

Diffusion TV

About

Since AI Generated video is taking off 🎉 I created Diffusion TV https://diffusiontv.vercel.app/ a mobile only project that helps you stay on top of the latest AI generated videos with a user experience inspired by Tiktok/Reels. Helps me to discover content, find my favorite artists, and keep up with the latest AI Generated videos.

This project was born out of a desire for me to stay on top of the cutting edge of AI video technology and to shine a light on the amazing work being done by many talented individuals. The best part is, it's open source so everyone can learn how to build this themselves.

My goal is to see if there is enough interest in this proof of concept to convert this from a PWA to a fully-fledged native iOS/Andriod application 📱

Note: The content in DiffusionTV based 100% on [Civitai's API](https://github.com/civitai/civitai/wiki/REST-API-Reference). None of the content is scraped from their website in order to remain compliant to their Terms of Service. All video content belong to their respective owners and creators.

Installation

Diffusion TV is a PWA or Progressive Web App. Meaning it will act and look like a native mobile application once saved to your home screen. Saving to home screen is simple:

Go to https://diffusiontv.vercel.app

iOS:

1. Open application in Safari

2. Select the Share icon

3. Select "Save to Home Screen"

Andriod:

1. Open application in Chrome

2. Open Settings

3. Select "Save to Home"

Features

The videos and engagement data comes from Civitai, however while Civitai is a great place to share models and images, video content is underserved. I wanted a way where I could follow my favorite creators, stay on top of the latest videos, and watch videos that aren't normally surfaced on Civitai by creating a unique recommendation algorithm based on clusters of video content using [CLIP](https://arxiv.org/pdf/2103.00020.pdf) as a baseline.

Recommendation Algorithm

video2cluster.excalidraw.png

Pasted image 20231106180947.png

Support the Project with Stars 🤩

I am running this project out of pocket at the moment. Costs are daily transcoding, storage and serving images and videos, and CDNs for this content. If you appreciate what I'm doing start me ⭐ on GitHub, https://github.com/shelbyt/diffusiontv send feedback and share with friends. Thank you!

[^1]: The next feature on the roadmap is to enable Civitai users to "claim" their profiles. This is challenging because Civitai has no "Login with Civitai". I have found a couple of ways around this to securely claim and verify ownership but it will take a couple weeks.

[^2]: Costs include daily jobs for transcoding, serving, and recommendation algorithms. Serving video is very expensive.

[^3]: Civitai's content filters are very aggressive and have many false positives. By loosening them I can double the video database but that also doubles costs.

[^4]: To reduce memory usage, the transcoder pipelines ensures that 90% of videos are <10MB. However, this can be reduced further and make for instant loading times.

[^5]: At the moment the video pipeline runs daily. It would be nice to run this more frequently and enable users who are not from Civitai to directly upload videos.

[^6]: PWA's are great but there are many limitations that I had to hack for it to work and be memory efficient. A native app would overcome many of these issues.

[^7]: [Civitai's API is broken ](https://github.com/civitai/civitai/issues/769]), so I fixed it.

Techstack

tl;dr NextJS front-end. Node and Python backend.

Open Source License

Diffusion TV is licensed under AGPL.
https://github.com/shelbyt/diffusiontv

---

Your support and feedback are what drive this project forward. Please star on GitHub, report issues or share your thoughts and ideas.

5

Comments