Find the Prompt Behind Any Old AI Image: Turn Your ComfyUI Output Folder Into a Prompt Library With mondo-harvest
The recipe for every image you ever made in ComfyUI is probably still sitting inside the file. A free tool reads it back out. The harder part is keeping that record alive.
Most people who make AI images have a folder they are slightly afraid of. Thousands of files with names like ComfyUI_04417_.png, a handful of them great, and no idea how any of the great ones were made.
The odd thing is that the answer is usually still there. ComfyUI, the free node-based tool many artists use to build image and video setups, writes the whole setup into the PNG images it saves by default: the prompt, the model, the settings, the random number that fixed the composition. It is like a cook who staples the recipe to the back of every plate. Almost nobody turns the plate over.
A small open tool released this week, mondo-harvest, does that for a whole folder at once. Point it at your outputs and it writes one file listing every distinct prompt you used, with the model, settings and files behind each one. It does not change your images. It only reads the hidden notes inside them.
That sounds like a filing chore. It is closer to finding your own lost sketchbook.
Why your old prompts are worth more than new ones
When you are making images cheaply, the expensive thing is not the image. It is remembering how you got the good one.
A prompt that worked is a tested piece of craft. It encodes the words that pushed a model toward the light you wanted, the lens feel, the texture. Lose it and you are back to guessing. Keep it, and the next job starts from your own best work instead of from a blank box.
Every artist eventually builds some kind of library for this: a notes file, a spreadsheet, a pile of screenshots. Those depend on you remembering to write things down at the moment you are most absorbed in the work, which is when nobody writes anything down. The files themselves never forget, as long as you do not wipe them.
What mondo-harvest reads
The tool reads two kinds of hidden notes.
ComfyUI's saved setup. A ComfyUI image carries the full node graph that made it. That graph can be messy: old prompts left in muted boxes, two prompt nodes where only one was connected, test text you forgot about. mondo-harvest follows the graph from the save step back to the step that actually drew the picture, so you get the prompt that made the image, not every text box on the canvas.
For each image it also records:
- the model file and text encoders used;
- any add-ons switched on, with their strengths (these are the small style or character files ComfyUI users call LoRAs; ones that were switched off are left out);
- the "avoid this" prompt, the sampler settings and the output size;
- the seed, which is the random starting number that, with the same settings, reproduces the same picture;
- whether a reference image fed the result;
- prompt text that sat in the setup but was not used, listed separately.
It handles PNG, WebP and JPEG images, and MP4, MOV, WebM and MKV videos saved by ComfyUI's video nodes.
Prompts other tools leave behind. Plenty of other generators and services write the prompt into standard image fields, such as captions or comments. mondo-harvest reads those too, but carefully. A field only counts when its name says it can hold a prompt and its text reads like one. Files marked as real photographs or rights-managed scans are skipped unless they also carry an AI label, because museum and stock captions look a lot like prompts.
What you get back
One file per folder, saved as harvests/<folder name>.json. JSON is a plain text format any code editor can open and any spreadsheet tool can import.
Each prompt appears once, however many images used it. Under it you get every setup it ran with, and under each setup, every file and seed. So if you made forty variations of the same idea over three weeks, you see one entry, with the dates of the first and last, and the forty seeds.
Two smaller details show the care in it.
Many add-ons need a trigger word at the start of the prompt, like "ohwx woman." Saved without the add-on, that word is noise. The tool spots trigger words that match the add-on's file name, takes them off the prompt and keeps them with the add-on instead, then lists everything it removed so you can check it guessed right. If you would rather keep trigger words in place, the --keep-triggers option turns this off (add-on tags written as <lora:...> are still removed).
Long crawls save progress every two minutes. If you stop a run of tens of thousands of files, running the same command again picks up where it left off.
Put this into practice this week
You need Python 3.10 or newer. Reading videos also needs ffprobe, which comes with the free FFmpeg tools. Nothing else.
1. Get the tool. It is a single file. Clone the repository (the README also points to a download of mondo_harvest.py from its releases page):
git clone https://github.com/hodgemann/mondo-harvest.git
cd mondo-harvest
2. Start with one small folder. Pick a subfolder of a few hundred images from a project you remember. You want to be able to tell whether the results make sense.
python3 mondo_harvest.py crawl /path/to/that/folder
3. Read the summary it prints. At the end it tells you how many files it read, how many had prompts, and which hidden fields it found. That last list is a quick health check on your folder. If most files report nothing, something upstream is stripping the notes (more on that below).
4. Open the harvest file and search it. Look for a word you remember from a prompt you liked. Copy the prompt, the model and the seed, rebuild the setup in ComfyUI and run it. If you get the same image, you have your recipe back.
5. Check the trigger list. Scroll to lora_triggers. If the tool removed a word that was really part of your prompt, rerun with --keep-triggers.
6. Now run the big folder. Once you trust the output, point it at the whole archive and let it run.
7. Make it a habit. Run a crawl at the end of each project. Keep the harvest files with the project. Over a year, that becomes the most useful reference you own.
The habits that destroy the record
The tool can only read what was saved. The real lesson here is about how files lose their recipe.
Re-saving through an editor. Open a PNG in an image editor, touch it up, export it, and the hidden notes may not survive. Keep the untouched original next to any edited version.
Uploading and downloading again. Some websites and chat apps strip hidden data from images on upload. The copy you download later can come back blank. Never let the uploaded copy become your archive copy.
Hosted services that write nothing. Some online generators leave no prompt in the file at all. For those, the only record is the one you keep yourself.
Renaming and converting in bulk. Batch converters that change format or shrink files can drop the notes along the way. Test one file before you run a converter on a folder.
The quickest test: run mondo-harvest on a folder you have edited and shared, and compare the "files with prompts" number to the total. The gap shows you how much of your history has already gone.
Where it falls short
It is new. The repository was created on November 11, 2026, reached version 1.3.0 the same day, has a couple of stars on GitHub, and is one person's project. That is fine for a tool that only reads files and never changes them, but expect rough edges.
It is a text file, not an app. There is no gallery view and no thumbnails. You search the harvest file in an editor or load it into something else. The author's own image app, Mondocraft, can import harvest files, but that app had not been released when this was written.
Trigger-word removal is a guess based on file names. It is careful, and it lists what it removed, but check it.
And it cannot bring back what is gone. If your best image was downloaded from a site that stripped it, mondo-harvest will politely report nothing.
The archive was there all along
The appeal of this tool is not cleverness. It is that it treats your output folder as a record of your craft rather than a pile of finished goods. The record was always there, stapled to the back of every image. Most of us just kept throwing it away.
Run it on one folder tonight. If it finds the prompt behind the image you have been trying to remake for months, you will start saving originals differently tomorrow.
- Sources: mondo-harvest on GitHub · Changelog · ComfyUI