Noise-DA Processing

Set expectations correctly before you touch this node: on a plain git clone of the pack, it will not run. It looks for a weight file - models/noise_da_denoise.pth (or _deblur.pth, _derain.pth depending on your mode) - inside the pack's own folder, and that file isn't in the repository and isn't fetched automatically. If you add this node, pick a model type, and hit run, you'll get a FileNotFoundError, not a picture. This isn't a bug you can work around with a setting; the weights genuinely aren't public, and the pack's own model-weights guide says as much: they're custom-trained and not distributed, so you'd need to train your own or ask the author directly.

What it's for, when you do have weights

Noise-DA - "Denoising as Adaptation" - is a real paper (Liao et al., accepted ICLR 2025, official code at github.com/KangLiao929/Noise-DA), and the idea is worth knowing even if you can't run this specific implementation yet: instead of training a restoration model directly on paired clean/noisy data, it frames denoising, deblurring, and deraining as a domain adaptation problem in noise space, closing the gap between synthetic and real-world degradation more directly than a standard supervised model does. The node wraps a RestorationNet (a U-Net-style encoder/decoder) for three tasks: denoise, deblur, derain.

Inputs and outputs that matter

• model_type - denoise, deblur, or derain, each looking for its own separate weight file.
• output_mode - residual (the model predicts a correction that gets added to the original, generally the more stable option and the tooltip's own recommendation) versus direct (the model's raw output is used as-is).
• strength and residual_scale - how strongly the correction gets applied and how it's scaled before blending; both only matter once the model is actually loading and producing output.
• blend_mode - normal, soft_light, or overlay blending between the original and the corrected image.

Outputs: processed_image and processing_info, a string documenting the settings used and citing the source paper.

Installing it

Through ComfyUI Manager, search Eric's Image Processing Nodes. By hand:

cd ComfyUI/custom_nodes
git clone https://github.com/EricRollei/Eric_Image_Processing_Nodes
cd Eric_Image_Processing_Nodes
pip install -r requirements.txt

Restart ComfyUI - the node itself will appear and load fine. It's running it that fails, because the required .pth file simply isn't part of the repository (it's gitignored on purpose, like all model weights in this pack) and there's no download URL wired up for it anywhere in the code, unlike NAFNet or the SwinIR architecture file. If you have your own Noise-DA checkpoint, or the pack author shares one with you directly, drop it at Eric_Image_Processing_Nodes/models/noise_da_<model_type>.pth and it'll load.

Where people get burned

Everyone hits the same wall first: a FileNotFoundError: Model file not found the instant they try to run this node, with no obvious next step in the UI. The fix isn't a pip install or a settings tweak - it's that the weights genuinely don't exist anywhere public yet. Don't spend time debugging your ComfyUI install over this; check the pack's Docs/MODEL_WEIGHTS.md and its GitHub issues for whether the author has since published a checkpoint, or use one of the pack's other AI denoisers (NAFNet or SCUNet) that either auto-download or at least have a known, documented weight source. If Noise-DA genuinely matters to your workflow, training your own from the official KangLiao929/Noise-DA repository is currently the only reliable path.