Auto Style Dataset

Dataset curation beats every knob in LoRA training - that's the oldest, most-replicated finding in the community, and the single biggest dataset sin is boring variety. If your style dataset's prompts all say "beautiful woman in a field," your style LoRA learns a field, not a style. Auto Style Dataset exists to kill that problem for you: it outputs 35 pre-written prompt strings spanning elderly fishermen, leaping ballet dancers, welding sparks, hummingbirds, sea turtles, cracked pocket watches, scattered chess pieces - a deliberate spread of subjects, scenes, and lighting so the style has to be the constant, not the subject.

Mechanically it's the simplest node in the pack. It reads 35 prompts from a config file (auto_style_dataset_prompts.txt, one per line) that ships inside the pack, and applies optional prepend and append text to every one. Thirty-five STRING outputs - prompt_00 through prompt_34 - each carrying one finished prompt. There are no required inputs, no models, no downloads. It's a text file with a node face.

How you'd actually use it

The prepend and append fields are where the power is. Prepend your trigger word and style anchor so every image carries it consistently ("in the style of STYLETRIGGER, "), append quality tags or a negative instruction. The KB's captioning rule for style training is the one that applies here: describe what you want to remain variable (the subjects already vary), and leave what should be fixed in the trigger or prepend. Wire the 35 outputs into a batch of empty latents with a shared seed, and you've got a 35-image style dataset in one queue - which is enough to train a passable style LoRA, per the small-dataset finding.

It's also genuinely useful outside training: as a fixed prompt bank for benchmarking, for testing a checkpoint's range, or for feeding a LoRA test prompt loop.

Inputs, outputs, gotchas

Just prepend and append (both optional, both plain strings), and the 35 prompt_* outputs. The text file lives at nodes/auto_style_dataset_prompts.txt inside the pack - you can edit it to make the 35 prompts your own, which is the intended customization path. If you change the file, restart or reload so the node re-reads it; the node loads it fresh per call, but ComfyUI caches imports.

One honest caveat: 35 fixed prompts are a starting point, not a complete dataset. Style training still wants varied lighting, environments, and composition across a few hundred images to be robust. This node gives you the curated seed batch with zero effort; the KB's broader guidance (diverse backgrounds over transparent ones, avoid heavy makeup if you're doing faces, caption what should vary) still applies to what you build around it.

Install

Part of TrentNodes: ComfyUI Manager (search "Trent Nodes"), or

cd ComfyUI/custom_nodes
git clone https://github.com/TrentHunter82/TrentNodes.git
cd TrentNodes && pip install -r requirements.txt

No extra dependencies for this node beyond the pack's core install. It's the kind of utility that looks trivial until you've hand-written 35 varied prompts once - then it stops looking trivial.