latentverse
Portal hopping with Stable Diffusion 👾
MIT License
Latentverse 🧑🚀
Latentverse uses latent consistency models (LCMs) and a custom LoRA (360Redmond) to generate 360-degree equirectangular photos in seconds.
It's like Rick's portal gun, or Google Street View for your imagination.
With the right set up, you can achieve near realtime generation. With an A100 and 4 steps, each inference takes ~1s. On an A10G (as shown in the video below), you get ~3s.
Getting started
There are two parts of the project, the Next.js frontend app and the backend server hosted on Baseten.
You can deploy the frontend web app locally as follows:
- Clone this repo.
-
cdintowebapp - Run
npm install
Unfortunately, you can't run the backend locally quite yet! We use Baseten, a serverless GPU provider for ML applications. The deployment folder has a Truss that can be deployed on Baseten.
- Create an account on Baseten.
- Run
pip install truss - Run
truss push deployment --publish. You may have to provide authentication credentials here. - Go to your app's dashboard on Baseten and grab the endpoint that has been set up.
- Update the endpoint on line 10 in
webapp/src/app/api/portal/route.ts.
The last step is creating a .env.local in webapp with the following Baseten credentials:
BASETEN_API_KEY=YOUR API KEY
BASETEN_ENDPOINT=YOUR API ENDPOINT (e.g. https://model-abcdef.api.baseten.co/production/predict)
You can now run npm run dev in webapp. By default, the web app should now be available at https://localhost:3000. Happy portal hopping!
Contributions and Licensing
LatentVerse was built by Varun Shenoy and David Song.
Feel free to provide any contributions you deem necessary or useful. This project is licensed under the MIT License.