EricRollei-Eric_Image_Processing_Nodes/external_tools/deepinv_service

DeepInv Service

This folder hosts the FastAPI microservice that exposes selected DeepInv models to ComfyUI nodes. The service is intended to run inside the dedicated virtual environment at A:\Comfy25\envs\DeepInv so it can install heavy dependencies without touching the primary ComfyUI venv.

Quick start

# Activate the DeepInv environment
& A:\Comfy25\envs\DeepInv\Scripts\Activate.ps1

# Launch the API
python external_tools\deepinv_service\service.py

The service listens on http://127.0.0.1:6112 by default. Override the port by setting the DEEPINV_SERVICE_PORT environment variable before launch.

API surface

• GET /health — quick liveness probe.
• POST /denoise — run a DeepInv denoiser on an image saved as a .npy tensor (NHWC, float32, range [0,1]). The request body must contain:
  • input_path: path to the tensor to process.
  • output_path: destination .npy file that will store the result.
  • model: one of drunet, dncnn, restormer, scunet, swinir.
  • sigma: optional noise level (defaults to 0.1).
  • prefer_gpu: optional flag to request CUDA execution.

Example payload:

{
  "input_path": "A:/tmp/input_image.npy",
  "output_path": "A:/tmp/output_image.npy",
  "model": "drunet",
  "sigma": 0.05,
  "prefer_gpu": true
}

Next steps

• Add endpoints for deblurring, super-resolution, and other recovery tasks using DeepInv's physics + reconstructor stack.
• Integrate an auto-starter from the ComfyUI launch batch so the service comes up automatically.
• Build ComfyUI nodes that call the API (e.g., via requests or asynchronous HTTP) and exchange tensors through the temporary .npy files.