D2 Save Caption

If you've ever run a WD14 tagger over a folder of training images and looked at the raw output, you know the problem: duplicates, quality tags you don't want in your LoRA, escaped parentheses, inconsistent blue eyes vs blue_hair. D2 Save Caption is the cleanup + write step for exactly that pipeline. It formats a tag string, filters it, and writes it to a caption file next to your image.

It's aimed at the dataset-prep workflow, and it pairs naturally with the rest of the D2 pack: D2 Folder Image Queue hands you image paths, D2 Load Text reads an existing caption, and this node writes the cleaned version back out. It's not a generation node - think of it as the "save as" step with opinions.

The inputs that matter

• base_filename - the source path your caption is named after. The node replaces the extension and saves there: d:/images/aaa.jpg → d:/images/aaa.txt. Feed it the image path from D2 Folder Image Queue.
• text - the caption body (from a tagger or D2 Load Text).
• extension - output extension (default txt).
• exclude_tags - tags to drop. Comma or newline separated. Prefix a line with regex/pattern/ and it excludes anything matching that regex. Matching ignores bracket escaping, so you can exclude rem_(re:zero) to kill rem_\(re:zero\) tags.
• prepend_tags - tags to force onto the front. Already-present tags aren't duplicated.
• replace_underscore - unifies the word separator in tags (the newer pack versions renamed this to a three-way word_separator option, so exact behavior depends on your version).
• trailing_comma - add a trailing comma (some training setups want it).
• ignore_case - case-insensitive matching for exclusion (on by default).
• backup - if a file with the same name exists, rename the old one to .bak before writing (on by default). This has saved me more than once.
• dry_run - preview mode: format and return the result without writing a file.

Outputs: text (the formatted caption) and file_path (the full save path, or the planned path under dry_run).

How it works

The pipeline is deterministic and documented: split → trim → unify word separators → strip bracket escapes → exclude → dedupe → prepend → trailing comma. Dedup keeps the first occurrence. If base_filename is empty, it stops with an error rather than save somewhere unintended - which is annoying until you realize it just protected you from writing 500 captions into a wrong directory.

Installing

Part of the D2-nodes-ComfyUI pack - install the pack, not the node.

• ComfyUI Manager → search "D2-nodes-ComfyUI" → Install → restart.
• Or cd ComfyUI/custom_nodes && git clone https://github.com/da2el-ai/D2-nodes-ComfyUI and restart.

Light dependencies (piexif, charset-normalizer), no model downloads. Pack gotcha: v32.0.0+ needs the ComfyUI V3 schema; older ComfyUI → pre-32.0.0 release.

Where people get burned

Run a dry_run before you let it touch real files - the text output shows you exactly what will be written, and it costs one queue. The exclude_tags regex syntax trips people up: it's regex/pattern/ (slash-wrapped), not bare. And remember backup only creates a .bak of the existing caption; if you've already got .bak files, they get overwritten. If you're captioning a few thousand images, build the exclude list with the pack's D2 Tag Report (which counts tag frequency in a folder) rather than guessing - it'll save you a full re-run.