DIGIT Sample Prompt Preset

When you train a LoRA, the sample prompts are how you judge the run - a fixed set of prompts regenerated each epoch so you can watch the model learn. DIGIT Sample Prompt Preset is the part of the DIGIT LoRA training suite that keeps those prompts consistent and reusable. Save a list of prompts as a named preset, load it back for the next training run, and every run of that style gets compared against the same gallery instead of whatever you happened to type that morning.

The lora-training essay in the modidex KB hammers home why this matters: training is evaluation, and evaluation is only meaningful if the test set doesn't drift between runs. This node is that test set, versioned.

The inputs

• action - the whole node is one action dropdown: save, load, list, delete, or preview (the default).
• prompts - the meat. One prompt per line, and they can carry the DIGIT placeholders: [trigger], [trigger_class], [trigger_phrase]. The default is a decent starter gallery - a portrait, an outdoor shot, a cinematic scene, a close-up, a city-street full body - all phrased around your trigger phrase.
• preset_name - the name to save under or load. Required for save/load/delete.
• description - a note so you remember why this gallery exists.
• trigger_word / trigger_class - used only by preview, to resolve the placeholders into concrete text.

Outputs: result (a status string for save/load/delete/list) and resolved_prompts (the prompts with placeholders filled in - the thing you'd actually wire into the trainer or a text encoder).

How it works

preview is the mode you'll use most while building a gallery: it takes the placeholder templates plus your trigger info and shows you exactly what the final prompts will read as before you commit them to a training run. save and load talk to the pack's local presets database, so presets survive restarts and live alongside the other training presets (trigger presets, naming presets, caption presets). list shows what you've got, delete cleans up.

Setup and usage

It's part of the DIGIT pack - ComfyUI Manager, search comfyui-digit, restart. No API keys, no models, works fully offline. The preset DB is local to wherever ComfyUI runs, so on a multi-machine fleet, presets don't follow you unless you sync the folder - a minor gotcha if you train on one box and render on another.

Typical flow: define your trigger word in DIGIT Trigger Preset, then here in Sample Prompt Preset build the gallery with [trigger_phrase] placeholders, hit preview to confirm it reads well, save it under a name like flux1_style, and load it every time you kick off a run of that style. It's a small node, but it's the difference between "oh, the model changed" and "oh, I changed the test" - and it makes sure it's the model.