Noise-DA Batch Processing

Read this before the single-image Noise-DA node's article if you found this one first: same model, same limitation. This is the batch-processing sibling of Noise-DA Processing, built to run a whole IMAGE batch through one loaded model instead of reinitializing per image. It shares the exact same requirement that trips up a fresh install - it needs a .pth checkpoint that isn't included in the repository and isn't downloaded automatically. On a stock install, this node throws a FileNotFoundError the moment you run it, batch or not.

What it's built for

Noise-DA - "Denoising as Adaptation," Liao et al., ICLR 2025 - treats restoration as a domain-adaptation problem in noise space rather than plain supervised denoising, and this batch node runs one of three trained variants (denoise, deblur, derain) across a stack of images with the model loaded once and reused, rather than reloading weights for every frame. That's the right shape for processing a folder's worth of scans or frames efficiently - assuming the weights are actually present, which for almost anyone cloning fresh from GitHub, they aren't.

Inputs and outputs that matter

• model_type - denoise, deblur, or derain. Each expects its own separate .pth file at models/noise_da_<model_type>.pth inside the pack folder.
• output_mode - residual (correction added to the original) or direct (raw model output).
• use_gpu - falls back to CPU automatically if CUDA isn't available.
• strength - blend strength between original and processed, applied per image in the batch.

Unlike the single-image version, this node has no residual_scale or blend_mode controls - it's the leaner batch path, trading some tuning granularity for throughput once weights are actually loading. Outputs: processed_images (the whole batch) and batch_info, a string summarizing the run.

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 loads cleanly; it's the model weights that are the blocker, and that blocker is identical to the single-image Noise-DA node - the pack's own model-weights documentation is upfront that these particular checkpoints are custom-trained and not publicly distributed, unlike NAFNet's or SCUNet's, which either auto-download or have a documented public source.

Where people get burned

The same wall as the single-image node, just hit while trying to process a whole folder: a FileNotFoundError on the first batch item, with the batch aborting entirely rather than partially completing. There's no partial-batch recovery here - if the model can't load, nothing in the batch gets processed.

Before spending time debugging this as an installation problem, confirm you actually have noise_da_denoise.pth (or the deblur/derain equivalent) sitting in Eric_Image_Processing_Nodes/models/. If you don't, this node can't do anything for you yet, and that's a weights-availability issue rather than something wrong with your ComfyUI setup. For batch restoration work that actually runs today, SCUNet Batch Processing or the batch path inside Comprehensive Method Comparison are the more reliable picks in this pack - both have either a documented weight source or a working fallback rather than a hard dependency on an unpublished checkpoint.