🤗 Dataset To LIST

The 🤗 dataset nodes talk to each other in an opaque HUGGINGFACE_DATASET type. The rest of ComfyUI - and especially the generic list/dict/string nodes of the Basic data handling pack - doesn't speak that language. 🤗 Dataset To LIST is the translator: it materializes your dataset into one plain Python list, which is the exact shape Basic's LIST → length, LIST → get item, and friends expect.

How it works

Three inputs, one output:

• dataset - anything the pack produced: the loader's dataset output or the result of a whole transform chain.
• column - leave it empty and each list item is a whole row dict; set it (e.g. text) and each item is that column's value. The "one column out as a list" mode is the one you'll use constantly - it's how you feed a corpus of review texts or prompts into string nodes.
• limit - cap how many rows get materialized (-1 = all). On a streaming dataset this is your download brake: To LIST (column=text, limit=100) pulls exactly 100 values off the lazy stream.

The output list is a LIST in the Basic-data-handling sense: one Python list value, delivered whole. That distinction matters because the pack ships two converters for two shapes - To LIST for whole-list nodes (length, get item, first, join), and 🤗 Dataset To Data List for per-row processing where Basic runs a node once per item. Pick based on what you're wiring into: flat operations want the LIST, row-wise transforms want the Data List.

A recipe that shows the point

Load (path=stanfordnlp/imdb) → 🤗 Dataset To LIST (column=text, limit=100) → Basic/LIST → length

That's "how many review texts do I actually have" in three nodes, no Python. Add a Filter before the converter and you're counting only the rows that passed the condition.

Gotchas

Materializing is the moment the graph stops being lazy - with limit = -1 on a non-streaming load you're already paying for the full download, and To LIST then builds the whole Python list in memory. For big data, set limit or put a Take first. Also keep column spelling exact: a typo means a missing-key error on the first row rather than a friendly warning. And be aware row values become plain Python - numpy scalars are converted, but image/audio objects pass through as-is, so a "list of images" is only useful to nodes that handle that type.

Installing it

Part of StableLlama/ComfyUI-huggingface_dataset. ComfyUI-Manager → search "Hugging Face dataset", or:

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

Restart ComfyUI. The Basic data handling pack is a separate install but well worth it - these converters are built around its list types.