ConvIR DeRain

Every other node in the ConvIR family asks you to pick a dataset and a model size. ConvIR DeRain doesn't. It has exactly one input - the image - and one job: take the rain streaks out. In a pack full of research-model wrappers this is the "it just works" node, which is also why it's the easiest one to recommend to someone who doesn't want to think about datasets.

What it does

It's the same ConvIR restoration backbone you'll find across this pack - a U-shaped convolutional encoder-decoder with redundant-feature pruning, from the Revitalizing Convolutional Network for Image Restoration paper (Cui et al., TPAMI) - trained to separate rain from scene. Feed it a rain-streaked frame and it returns a clean, same-size version. It's deterministic and fast-ish: one forward pass per frame, and it handles batches, so a video loaded as an IMAGE stack works with a ComfyUI progress bar.

The single input

• src_img - that's it. The node hardcodes the deraining model (the large size of the deraining checkpoint), so there's nothing to configure and nothing to get wrong.

The res_img output is the de-rained IMAGE at the same resolution, ready for a Save node or to feed an upscaler.

Installing it

The pack install is the standard one - ComfyUI Manager, search FM_nodes, or:

cd ComfyUI/custom_nodes
git clone https://github.com/FuouM/FM_nodes

then restart. Requirements are just torch and einops; no keys, no services, nothing exotic. The model, as always with this pack, is manual: grab deraining.pkl from the ConvIR Google Drive and put it directly in custom_nodes/FM_nodes/models/convir/ - the root of the convir folder, not a subfolder. The code looks for models/convir/deraining.pkl, so if you queue and get a file-not-found, check whether it ended up in a deraining/ subfolder by accident.

Where people get burned

Honestly, very little to trip on here - that's the appeal. The one expectation to manage is that removing rain streaks is not the same as removing rain distortion: heavy rain with big drops and water on the lens will leave soft patches where the model couldn't reconstruct the scene, and the output can look slightly smoothed in those spots. Light-to-moderate streaks vanish cleanly, which is the common case. It's the node you add to a video pipeline and then forget exists until it saves a clip you'd otherwise have to reshoot.