ComfyUI Invertible Rescale (Downscale only)

A single-node ComfyUI wrapper for Invertible Image Rescaling (IRN) downscaling with:

• Auto-download of models (selected or all) from the official Google Drive folder
• Automatic discovery of local .pt/.pth models
• BasicSR .pth loader (vendored minimal IRN network) for IRN_x2/x3/x4
• Device and precision control
• Tiling for low VRAM
• Safe handling of sizes not divisible by the scale
• Alpha channel preservation
• Built-in bicubic/area fallback

Install

1. Copy this folder to:

   ComfyUI/custom_nodes/comfyui_invertible_rescale
   

2. Restart ComfyUI.

Optional dependencies:

pip install torch torchvision
pip install gdown   # for Google Drive auto-download

Models

Official IRN pretrained models (IRN_x2.pth, IRN_x3.pth, IRN_x4.pth) are hosted at:

• https://drive.google.com/drive/folders/1ym6DvYNQegDrOy_4z733HxrULa1XIN92

This node will:

• Auto-download the selected model if auto_download_selected is ON and missing.
• Optionally download all available models if auto_download_all is ON.
• Discover and list all .pt/.pth files in:

  ~/ComfyUI/models/invertible_rescale
  

You can also set direct URLs (and SHA256) in model_registry.json if you prefer non-GDrive hosting.

Node: IRN Downscale

Inputs:

• image: an IMAGE tensor
• model_key:
  • auto_best_match: uses IRN_x{scale} if found (downloads if enabled), else builtin.
  • builtin_bicubic: standard PyTorch downscale (bicubic/area).
  • Any discovered local/registered model (dropdown).
• scale: 2 / 3 / 4
• device: auto / cuda / cpu / mps
• precision: fp16 / bf16 / fp32
• tile_size, tile_overlap: enable tiling to reduce memory load
• not_divisible_mode: pad / crop / resize_to_multiple
• builtin_mode: bicubic / area (only for builtin)
• keep_alpha: preserve alpha (downscaled with area)
• auto_download_selected, auto_download_all, use_google_drive, models_dir_override

Outputs:

• image: downscaled image
• meta: model used, device, timing, available models, and any load error

License

MIT