Illustrious Open Prepared Dataset (After Review)

IllustriousDatasetOpen is the least glamorous node in the Illustrious LoRA Tools pack and the one you'll be grateful for. It does one thing: takes the name of a dataset run you already prepared, tagged and eyeballed, and hands it back as a dataset handle so you can go straight to training.

Why that matters is about what the alternative costs you. The pack's full pipeline re-imports and re-tags on every queue - that's four nodes of work, and if you've hand-edited captions the auto-tagger will not preserve them unless you remember to leave overwrite_prepared_captions off. The "reviewed" path exists so you tag once, fix once, and then train as many times as you like against that frozen copy. That's the workflow the pack's own 02_Advanced_Reviewed_Training example ships, and it's the one you want once your captions are done.

How it works

There's not much to it, which is the point. The node builds a handle containing the run name you typed, then has the shared dataset_manifest helper open ComfyUI/output/illustrious_native/<run_name>/dataset.json and validate it. That validation is stricter than you'd expect: the manifest has to be version 1, has to describe between 1 and 1000 images, has to be under 4 MB, and every image file it lists has to still exist on disk with a supported extension. Fail any of those and you get a plain error instead of a half-loaded dataset.

The node also declares itself always-dirty, so it re-reads the manifest on every queue rather than serving a cached handle. That's deliberate and it's what makes the edit loop work: fix a caption, queue the graph again, and the trainer sees the corrected file.

The one input, the two outputs

run_name is a string, and its default value is literally PASTE_PREPARED_RUN_FOLDER_NAME - the author's way of telling you not to guess. It wants the folder name, not a path. Something like dataset_a91f3c04be (see below), not ComfyUI/output/illustrious_native/dataset_a91f3c04be and definitely not an absolute path. Path-like input is rejected by the shared path guard, which refuses anything absolute or containing ...

Where do you get that name? From upstream. IllustriousDatasetPreview reports it as run when you click the node's text box, and the prepare/crop/edit nodes all echo theirs in their own reports. Prepared runs are named <label>_<hex>, so the suffix is the part that matters - the label alone won't resolve.

Output dataset wires into IllustriousBatchAutoTag, IllustriousCaptionAudit, IllustriousCaptionEdit, IllustriousDatasetCrop, IllustriousSDXLLoRATrain - anything that expects an ILLUSTRIOUS_DATASET. Output report is the manifest as JSON: the image count, the per-file list with dimensions, and the status field, which reads prepared, tagged, or a failure state. The trainer checks that status as part of its pre-flight, so it's worth a glance before you start a 20-epoch run.

Install

Same as the rest of the pack - Manager (search ComfyUI-IllustriousLoRATools, shown as Illustrious LoRA Tools) or:

cd ComfyUI/custom_nodes
git clone https://github.com/katorikonoe-ai/ComfyUI-IllustriousLoRATools
pip install -r ComfyUI-IllustriousLoRATools/requirements.txt

Use ComfyUI's own Python for the requirements so your CUDA PyTorch build survives, keep native/, web/ and crop_routes.py next to __init__.py, and restart the server. This node itself only needs Python's standard library plus whatever the pack already imports - the heavy lifting is in the training path, not here.

Where people get burned

The name-versus-path confusion is the number-one trip here, and the error message is honest about it: Use a relative path under the configured root, with / separators means you gave it a path when it wanted a bare folder name.

Second: prepared runs live in ComfyUI/output/. If you wiped your output folder, cleared it with a cleanup script, or moved to a fresh ComfyUI install, the run is gone and nothing will find it. There's no registry, no database - just directories on disk. Third: because this node skips IllustriousDatasetPrepare, it also skips the import and its limits. A run with more than 1000 images, or one whose manifest has been hand-edited into a shape the validator doesn't like, is rejected here rather than at train time. That's a feature; you find out in two seconds instead of twenty minutes into a queue.