Save Prompt

What it is

Save Prompt writes a text string to disk as a file - but the format and the placeholders are built for one very specific, very common job: batch captioning for training datasets. If you've ever run a captioner over a folder of images and needed each caption saved next to its image as a .txt sidecar, this is the node that does it inside ComfyUI instead of in a script.

The reason this pattern exists is the dataset convention: a LoRA training set is pairs of files that share a basename - 0001.png and 0001.txt. The trainer walks the folder and reads the caption from the same-name text file. This node exists to produce those pairs, with the source filename baked into the output filename automatically.

How it works

It takes your text, builds a filename from placeholders, and writes the file in one of three formats. txt is the plain caption (and the default - the one trainers want). csv writes name,prompt,negative_prompt rows, useful for prompt-styler style collections. json writes structured output and adds an optional nsfw_level tag (auto-detected from text keywords, or disabled). A JSON-only extra: batch source files keyed by their full filename.

The write modes matter for how you batch:

• new - numbered files (prefix_0001.txt, prefix_0002.txt), the safe default for separate outputs.
• overwrite - clobber the file each run.
• append - add text to an existing file, opening and closing each time.
• append_batch - keeps the file handle open across a batch for speed, auto-closing after 10 seconds idle. This is the one for captioning hundreds of images into one file.
• keep - skip if the file already exists, for resumable runs.

The placeholder engine is the whole point. %source_filename substitutes the source image's basename, %source_folder and %source_base_folder its folder path, plus %counter, %date, %time. The default filename_prefix is just %source_filename, so a caption for D:/images/cat.png lands at cat.txt in the same folder - exactly the sidecar convention. You get the source filename via the optional filename_opt string, or more powerfully via pipe_opt from a Load Image From Folder pipe, which also enables the folder placeholders.

The inputs that matter

• text - the caption/prompt to save.
• use_source_folder - save beside the source image, ignoring output_path. Default on.
• filename_prefix - the placeholder template, default %source_filename.
• extension - txt / csv / json.
• write_mode - new / overwrite / append / append_batch / keep.
• filename_opt / pipe_opt - how the node learns the source filename; pipe_opt wins if both are connected.

Output: text - the saved text, passed through as a list.

When you'd use it

Batch captioning pipelines feeding a training run, saving styled prompts into a CSV for reuse, and any workflow where the caption must land in the same folder as its image with a matching name. If you're building a dataset by hand in ComfyUI, this is the bridge between "I generated the images" and "the trainer has its captions".

Installing it

It's part of ComfyUI_Eclipse:

cd ComfyUI/custom_nodes
git clone https://github.com/r-vage/ComfyUI_Eclipse
pip install -r ComfyUI_Eclipse/requirements.txt

then restart, or install via Manager (search "Eclipse"). The usual lineage note applies: formerly RvTools_v2, legacy nodes removed in v4.0.0, bundled migration tool for old graphs.

Common issues

The big one is placeholders resolving to nothing: if neither filename_opt nor pipe_opt is connected, %source_filename has nothing to substitute and your files land with literal placeholder text in the name. Connect a source, or use a plain prefix. And if append_batch files feel like they close unexpectedly, that's the 10-second idle auto-close working as designed - reopen on the next call. For trainers specifically: make sure your trainer actually wants a .txt sidecar (kohya and ai-toolkit do); the convention is same basename, one caption per file.