Load Image (from Folder)

Load Image picks one file. Load Image (from Folder) picks a whole folder and hands you every image in it as a list. That single difference is what makes it a dataset node rather than a file node, and it's the node you reach for when "one image at a time" isn't the shape of your problem - you want all of them, right now, in your graph.

What it is

One input: folder, a dropdown of subfolders inside your input directory (ComfyUI/input). One output: images, a list of image tensors - one per file. Supported formats are PNG, JPG, JPEG, WEBP. The node scans the folder, loads every matching file in order, and returns them as a list where each image stays its own entry rather than being fused into one batched tensor.

The list output is the detail that matters. A plain batched tensor would be a single [N, H, W, 3] block - which falls apart the moment your images have different dimensions, or when you need to address them individually. A list keeps each image separate, which is what lets you zip it against a parallel list (captions, prompts), shuffle it, or feed each entry through a per-item pipeline. If you've used its sibling Load Image-Text (from Folder), the mechanism is the same; this one just drops the text side and returns images alone.

Where it fits

Two main jobs:

• Batch processing. "I have 200 images in input/batch/ and I want them all upscaled/face-detailered/converted" - this node plus a list-aware loop is the built-in answer. The community's standard alternative is WAS Suite's batch loader with an incrementing counter, but this node loads the whole set at once instead of one-per-queue.
• Training data. This is a dataset node in the literal sense - it reads the same folder-of-images structure the trainers use. Pair it with the image-text loader and the dataset save nodes, and ComfyUI covers the full data loop in core.

A practical note for training-adjacent work: the folder it reads from is ComfyUI/input, which is also where Load Image uploads land. Keep your working datasets in a clearly named subfolder - input/dataset/ - so you don't accidentally batch-process your entire loose pile of test images.

Getting it and where it goes wrong

It ships with ComfyUI core, a newer experimental addition in comfy_extras/nodes_dataset.py - no install, no model files. Gotchas to keep straight:

1. It needs a subfolder. The dropdown lists input subfolders; files dumped loose in input/ itself aren't the target here. The node's job is folder-scale loading.
2. Mixed dimensions are fine - for a list. Because it's a list output, images of different sizes load happily; the pain only starts if you feed that list into a node expecting a single batched tensor. If a downstream node chokes, the fix is batching/resizing first, not blaming the loader.
3. It's experimental. The list-based API is the current shape, and experimental nodes can reshape as ComfyUI's new node system matures.

If you've ever wanted to throw a folder at the graph and walk away, this is the built-in way to do it.