DataLoader Preview (Output)

ComfyDL ships two DataLoader preview nodes that look almost identical, and the difference is one flag that changes everything about how you use them. CdlDataLoaderPreview (Output) is the convenience version: it's registered as an output node, which means ComfyUI displays its rendered grid directly in the UI and always executes it - you don't have to remember to route its IMAGE into a separate preview node. For quick inspection - "let me just see what this dataset looks like" - this is the one you want.

The plain DataLoader Preview node next to it does the same rendering but behaves like any other node: it emits an IMAGE you must wire somewhere yourself. This variant exists so the grid just appears, which is what a beginner reaching for "preview" almost certainly means.

How it works

Under the hood it's a thin wrapper: the node's execution is literally delegated to the plain preview's logic, then the result is returned with the output-node flag set. So the mechanism is identical - pull one batch with next(iter(dataloader)), figure out whether the batch carries labels, adapt to the tensor layout (RGB permute, normalization to [0, 1], grayscale for single-channel), render a num_rows × num_cols matplotlib grid capped at max_samples, and rasterize it back to an IMAGE.

The rendering uses matplotlib with the headless Agg backend, so it works fine on a server without a display. The one behavioral difference worth internalizing: as an output node it runs even when nothing downstream is consuming it - which is exactly what you want for a "show me the data" node, and slightly wasteful if you wire it into a big graph and forget it's there.

Inputs and outputs that matter

• dataloader - any cdlDataloader (Fashion-MNIST, Bananas Detection, Load Array output, …).
• num_rows (default 2), num_cols (default 4) - grid dimensions.
• max_samples (optional, default 32) - cap on how many samples render.

The image output is still a real IMAGE tensor, so you can both let it display in the UI and wire it onward into a save or image-processing node if you want a copy.

Installing ComfyDL

cd ComfyUI/custom_nodes
git clone https://github.com/Cynthia-lxx/ComfyDL ./ComfyDL
pip install -r ./ComfyDL/requirements.txt

Restart ComfyUI and look under Datasets. ComfyDL's dependencies are light (matplotlib, IPython, matplotlib-inline), no downloads here. If ComfyUI Manager can't find "ComfyDL", clone - the pack is young.

Gotchas

Same first-batch caveat as its sibling: it previews batch zero only, so a non-shuffled dataset can show you an unrepresentative slice. And know that as an output node it executes on every run whether you're looking at it or not - leave one of these in a graph you're iterating on heavily and it's a small amount of matplotlib work happening every queue. For a one-off "show me the data" check, though, this is the friendlier half of the pair.