遮罩数据集路径

Normal LoRA training treats every pixel of your training images as equally valuable. Sometimes that's wrong. If your images have busy backgrounds, watermarks, or a subject that's half out of frame, you want the trainer to ignore all that and focus on the thing you actually care about. That's what MaskDatasetPathNode is for: it pairs your training images with a matching folder of masks, so the loss only counts what you've left white.

This is a relatively recent addition to the pack (the README logs "support mask training" in the Nov 2025 update), and it plugs straight into the same dataset pipeline as the plain GeneralDatasetPathNode.

How it works

You give it two folders: dataset_path, full of training images, and mask_path, full of masks. The node checks both folders exist (it prints a warning to the console if either is missing) and returns an input_path config dict - {"path": ..., "mask_path": ...} - that you wire into a GeneralDatasetConfig node, exactly like any other dataset path node.

The naming rule matters: each mask file must match its image's filename. image_001.jpg pairs with image_001.png (or whatever the mask extension is). If the names don't line up, the trainer can't pair them and your masking silently does nothing.

The mask semantics come straight from the author's tooltip: white = train, black = masked out, gray = train partially. So a pure subject-isolation mask is mostly white with black background. The gray middle ground is the interesting bit - it's effectively a soft weight on how strongly those pixels count toward the loss, which is useful when your subject edge is soft or you want the background to barely register rather than disappear entirely.

The inputs that matter

Only two, both required:

• dataset_path - the folder of training images.
• mask_path - the folder of mask images, same filenames as the training images.

Output: input_path, into GeneralDatasetConfig.

Install

Standard for this pack - ComfyUI Manager (search "Diffusion_pipe_in_ComfyUI") or:

cd ComfyUI/custom_nodes
git clone --recurse-submodules https://github.com/TianDongL/Diffusion_pipe_in_ComfyUI.git
cd Diffusion_pipe_in_ComfyUI
git submodule init && git submodule update
pip install -r requirements.txt

Linux/WSL2 only - on plain Windows this pack won't train at all.

Where people get burned

Three classics. First, filename mismatches: the mask system is filename-driven, so if you batch-renamed your images and not your masks, you'll train on unmasked data and never know. Second, missing the gray semantics - people paint all-black backgrounds and wonder why the subject edges look cut. A little gray feathering at the boundary trains smoother than a hard black/white edge. Third, pointing the mask path at the wrong folder entirely; the node warns in the console but happily returns a config anyway, so skim the logs on your first run. Masked training is genuinely one of the strongest tools for fixing a dataset with noisy backgrounds - it's worth doing the masking right.