DIGIT Dataset Manager

The DIGIT Dataset Manager is the first step of the pack's LoRA training pipeline, and it's the one that stops you from training on a mess. It does the dataset housekeeping in four actions - scan, create, validate, and stats - so your training folder is a known quantity before a single caption is written or a single step is trained.

Here's the workflow-shaped reason this node exists: training data quality beats every training knob, and the most common dataset failure is just sloppy plumbing - a stray 128px image mixed into a folder of 1024px ones, a missing caption, a source folder you forgot to copy. This node is the plumbing check.

How it works

The four actions:

• scan (default) - look at dataset_path and report what's actually there: image count, resolution info, caption coverage. The "am I about to train on junk?" check.
• create - build a new dataset by copying images from source_path into a new dataset folder, optionally filtering out anything below min_resolution (default 512). That filter is the point - a LoRA trained on tiny images learns to make tiny images.
• validate - check an existing dataset for problems: missing captions, wrong formats, undersized images.
• stats - summary numbers for the dataset.

Inputs worth knowing: dataset_name (default my_dataset) names a create target, caption_ext (.txt) says what a caption file looks like, copy_images (default true) controls whether create copies the actual files or just references them, and min_resolution is the quality gate. Outputs are dataset_path (where the dataset ended up - wire this into the Captioner), report (the human-readable findings), and image_count.

Installing it

Standard pack install:

cd ComfyUI/custom_nodes
git clone https://github.com/thedepartmentofexternalservices/comfyui-digit.git
cd comfyui-digit
pip install -r requirements.txt

Or ComfyUI Manager → search comfyui-digit → install → restart. This node is fully local - no cloud credentials - and it pairs with the rest of the training suite: the Dataset Manager feeds dataset_path into the DIGIT Captioner, which feeds into Dataset Prep or straight to the trainer.

How it fits the real workflow

The habit that pays off: run scan before you caption and validate after. Scan catches the 200px screengrab hiding in your character set; validate catches the image whose caption got written to the wrong filename. Both are five-second runs that catch the failure modes that show up later as "the model doesn't look like my character." And when you create a dataset, remember the resolution filter is a floor, not a suggestion - character LoRAs want clean high-res sources with diverse angles, backgrounds, and lighting, and this node is where you enforce that before it costs you a training run.