DIGIT Captioner

Where DIGIT Batch Caption is a standalone "caption this folder" tool, the DIGIT Captioner is the version that lives inside the pack's LoRA training suite. Same underlying idea - Gemini writes captions for your dataset through Vertex AI - but organized around training runs: it only captions what's missing, can re-caption everything, handles a single image, and reads its captioning recipe from a saved preset.

You'll reach for this when your workflow is already built on the DIGIT training nodes and you want captioning to behave like a step in the pipeline rather than a one-off job. The default action is caption_uncaptioned, which is the right instinct: caption only the images that don't have a sidecar yet, and never touch the ones you already captioned by hand. Re-running is cheap because it does nothing.

How it works

It scans dataset_path for images, checks for existing caption files (caption_ext, default .txt), and sends each image to Gemini with your system prompt and prompt template. The actions tell it which subset to work on:

• caption_uncaptioned (default) - only images missing a caption.
• caption_all - every image, respecting overwrite.
• caption_single - one image via single_image_path, for spot-fixing.
• recaption_all - force everything, ignoring existing captions.
• preview - show what it would write without committing.

The inputs that shape output: system_prompt (a solid detailed-captioning default ships with it), prompt_template, model (gemini-2.5-flash default), temperature (0.4 - keep it low, captioning wants consistency not creativity), max_tokens (300), and overwrite. The caption_preset field is where you load a saved recipe from the DIGIT Caption Preset Manager, which is the neat part: tune a preset once, reuse it across datasets instead of re-typing system prompts.

Outputs: report (a summary of what happened), last_caption (the most recent caption written, handy for spot-checks), and captioned_count.

Installing it

Part of digit-comfyui:

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. The full training suite wants the extra deps too (pip install -r requirements-training.txt) if you're going all the way to the trainer, but the Captioner itself only needs the base requirements plus your GCP auth: gcloud auth application-default login and a Vertex AI–enabled project, auto-detected or set via gcp_project_id.

Notes from the trenches

Two things worth knowing from how people actually caption. First, natural-language captions from Gemini are the right format for the modern LLM-encoder bases (Flux, Qwen-Image, Z-Image) - write what should vary, leave identity features undescribed, and this node's detailed default prompt is already pointed that way. Second, don't blow up temperature; a 0.4 default exists because a captioner that gets creative is a captioner that invents details your model will bake in. And since this is a hosted captioner, budget for per-call GCP billing on big datasets - caption_uncaptioned is also your friend there, because re-runs genuinely cost nothing.