Remote Upscale for ComfyUI

Offload image upscaling to a remote server with a dedicated GPU. Useful when your main ComfyUI machine has limited VRAM or when you want to use a different GPU for upscaling tasks.

Features

• Remote upscaling via HTTP API
• Real-time progress bar in ComfyUI
• Support for .pth, .pt, and .safetensors upscale models
• Configurable tile size for memory management
• Model caching on server for faster subsequent runs

Installation

ComfyUI Node (Client)

1. Copy the remote_upscale folder to your ComfyUI custom_nodes directory
2. Restart ComfyUI
3. The "Remote Upscale" node will appear in the "image/upscaling" category

Upscale Server

1. Copy the server folder to your remote machine
2. Install dependencies:

   pip install -r requirements.txt
   

3. Install PyTorch with CUDA/ROCm support appropriate for your GPU
4. Set the models path and run:

   export UPSCALE_MODELS_PATH="/path/to/your/upscale_models"
   python server.py
   

The server will start on port 8189 by default.

Usage

1. Start the upscale server on your remote machine
2. In ComfyUI, add the "Remote Upscale" node
3. Set the server_url to your server address (e.g., http://192.168.1.100:8189)
4. Select a model from the dropdown
5. Connect an image input and run the workflow

Node Parameters

• server_url: URL of the remote upscale server
• model_name: Upscale model to use (fetched from server)
• tile_size: Tile size for processing large images (default: 512)
• show_progress: Enable progress bar during upscaling

Refresh Models

Use the "Refresh Remote Models" node to update the model list from the server.

Server API

| Endpoint | Method | Description | |----------|--------|-------------| | /models | GET | List available models | | /upscale | POST | Upscale image (no progress) | | /upscale_stream | POST | Upscale with SSE progress | | /result/{id} | GET | Fetch result after streaming | | /health | GET | Server health check |

Requirements

Client (ComfyUI)

• requests

Server

• fastapi
• uvicorn
• python-multipart
• torch (with CUDA or ROCm)
• torchvision
• spandrel
• pillow
• numpy

License

MIT