Clarity-Upscaler

This repository contains the code for the Clarity-Upscaler deployment. The initial codebase was taken and modified from https://github.com/philz1337x/clarity-upscaler/tree/main.

• See PR: Ar-289-deploy

Deploying the Clarity-Upscaler Model on Replicate

1. Install replicate-cog on a local GPU machine:

   sudo curl -o /usr/local/bin/cog -L "https://github.com/replicate/cog/releases/latest/download/cog_$(uname -s)_$(uname -m)"
   sudo chmod +x /usr/local/bin/cog

   For the latest installation instructions, see replicate-cog.

   Note: While cog technically allows deployment from non-GPU machines, deploying this pipeline from a non-GPU machine (e.g., Macbook M3 Pro Max) has been unsuccessful. We recommend using either a personal GPU machine or a Lambdalabs GPU instance (A100 40GB) for making changes and pushing the model.

2. Clone this repository:

   git clone https://github.com/sdtechdev/replicate-clarity-upscaler
   cd replicate-clarity-upscaler

3. Download the required models and checkpoints:

   python download_weights.py

   To test if the pipeline is working properly, make a prediction locally:

   sudo cog predict -i image=@init.png

4. Either create a new model on the sdtechdev Replicate account or use the existing model repository: clarity-upscaler-4-deployment.

5. Push the model to Replicate:

   # First, login to Replicate
   sudo cog login
   
   # Then push the model
   sudo cog push r8.im/sdtechdev/clarity-upscaler-4-deployment

   Building and pushing the Docker image (approximately 27GB) can take 10-12 minutes depending on available cache images and internet speed.

Note: A GitHub Actions workflow is available to automate model deployment to Replicate (see .github/workflows/push.yml). However, due to the large Docker image size, the workflow fails on default GitHub runners. A custom-hosted runner is required to successfully push using GitHub Actions.