文件夹图像载入器

ComfyUI's stock LoadImage has you pick a file from a dropdown that only shows your input folder - and it caches that list until you refresh. That's fine for a human clicking around. It's a wall if your workflow is a batch job that processes whatever landed in a folder since the last run. FolderImageLoaderNode (display name "Load Images from Folder") reads a directory at runtime: you give it a path, it grabs every .png in there, and hands you the images and their filenames as two lists. No refresh button, no dropdown, no restart.

It's the entry point of the whole ComfyForEach pack, which is built around the pattern of "folder per task, process each image one at a time." You'll almost always pair it with the pack's two selectors: index an image out of image_list, index its name out of image_name_list in lockstep, and you've got a per-file batch loop.

How it works

Mechanically it's honest, short Python. It joins root_path/task_id/suffix_path, checks the folder exists (and throws a clear ValueError if it doesn't), lists .png files sorted alphabetically, then for each one: opens with PIL, applies exif_transpose so phone photos aren't sideways, converts 16-bit mode 'I' images down to 8-bit range, converts to RGB, normalizes to 0–1 floats, and wraps each image in its own one-frame tensor.

That last part is the one thing to internalize. The node returns a Python list of tensors - one tensor per file, each shaped (1, H, W, 3) - not a single batched IMAGE. The outputs are named image_list and image_name_list on purpose. Don't plug image_list straight into a KSampler and wonder why the frontend frowns at you; index it first with IndexedImageSelectorNode.

Inputs that matter

Three strings, and honestly only one is interesting:

• root_path - base directory, default empty. Set this to your images folder.
• task_id - the middle path segment (default abc123). This is where the "folder per task" idea lives: point one ComfyUI instance at a root, and each task's images live in root/task_id/....
• suffix_path - optional extra subdirectory; the pack's pattern is leaving it blank.

Outputs: image_list (IMAGE, one tensor per file) and image_name_list (STRING, the filenames in the same order). The name list is your map back to source files after processing.

Install

ComfyUI Manager, search ComfyForEach, or:

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

Restart and you're done. Dependencies are pillow, boto3, opencv-python-headless, numpy, and torch (ComfyUI's own). No model files to fetch. The README's instructions for patching ComfyUI's execution.py etc. are only needed for the AWS EventBridge error reporting - skip them for plain batch loading.

Where people get burned

• Folder not found - it raises immediately with the full path in the message. Double-check your root_path is relative to where you launched ComfyUI.
• PNG only - it silently ignores .jpg, .webp, anything not ending in .png. Feed it a mixed folder and it just sees fewer files. A lot of users hit this and assume the node is broken.
• The list-vs-batch thing again: if you're coming from a pack like rgthree or Impact Pack that flattens lists for you, remember this one expects you to do the indexing yourself. That's the pack's model, and it's fine - it just means the selector nodes aren't optional.

This is a no-frills utility with real value for scripted, headless runs - the kind of thing the same author built the Rabbit-Hole pipeline tool around. For an interactive workflow you may not need it. For "process whatever showed up in this folder," it's exactly right.