Your Old Midjourney Images Still Have Their Prompts

If you have been using Midjourney for a while, you have a folder of images whose prompts you can no longer remember. Most of those prompts are not gone. Midjourney wrote them into the images themselves — and, as a backup, into the filenames.

Midjourney does not embed a workflow graph the way ComfyUI does. But it does record your prompt. A current Midjourney PNG download carries the prompt as text in the image’s Description metadata field, with your parameters and a Job ID appended. The download’s filename is built from the prompt text too.

That gives every Midjourney image two copies of the same record. The trouble is that both are fragile, in different ways — and neither is much use until something reads it.

Two Places Midjourney Writes Your Prompt

Inside the image. The Description field holds the prompt you wrote, followed by your parameters and a Job ID, run together as one block of text. The prompt is all there, but it does not look like a prompt. Until something separates the parts, it reads as noise, which is why so many embedded prompts sit in libraries unseen and unsearchable. We have written up exactly which fields a Midjourney export carries.

In the filename. Midjourney names downloads after the prompt, alongside a job identifier. Its naming format has changed over the years, so a long-running archive contains more than one pattern.

Each copy fails differently. The embedded copy can disappear: re-saving an image, converting it, or passing it through a service that strips metadata will remove it. The filename copy can be renamed away — and, for long prompts, it was never complete in the first place.

What Numonic Now Does

Numonic now recovers prompts for images already in your library, not only for new imports, and it reads both copies:

• The embedded prompt is surfaced cleanly. Your parameters and the Job ID are separated out, so the prompt field holds the prompt you actually wrote rather than a string with an identifier stapled to the end.
• The filename is the fallback. Where an image’s embedded metadata is gone, the prompt is read from the filename instead — in both the current and the older Midjourney naming formats.

Once recovered, prompts behave like any other prompt in the library. They are searchable. You can find every image you made that mentioned a particular subject, style, or phrase, across years of work, without having remembered to record any of it at the time.

What Comes Back Is What Survived

One limit is worth stating plainly, because it affects the prompts you care most about.

The filename copy is a shortened copy. Filenames have a length ceiling, and Midjourney writes the prompt into the filename only up to that ceiling. Anything beyond it is cut off at download time — before the file ever reaches your drive. A short prompt survives in the filename whole. A long one arrives already truncated, usually mid-phrase.

The download we checked makes the point. Its prompt ran to 835 characters of description, followed by three parameters. The filename kept the first 55 characters, stopped part-way through a word, and carried none of the parameters. The embedded Description field had every character and every parameter.

The cruelty of this is that it is inversely targeted. Throwaway prompts survive intact. The elaborate ones — stacked style references, long clauses of negative direction, the parameter tails you spent an afternoon tuning — are exactly the ones most likely to be clipped.

This only bites when the filename is the last copy standing. If an image still carries its embedded Description, the full prompt is there and the truncated filename does not matter. When the embedded copy is gone, the filename gives you the opening of the prompt: often enough to identify the image and search for it, but not always enough to run it again. Treat that as a record rather than a reproduction. Where it stops, it stops for good.

Why This Matters More Than It Sounds

There is a version of this feature that sounds like housekeeping. It isn’t, for a reason worth spelling out.

A prompt is the closest thing generative work has to a sketch. It is the record of intent — what you were trying to do, the vocabulary you had settled on, the phrasing that turned out to work. Practitioners develop their own language over time, and that language is spread thinly across thousands of individual generations.

When prompts are unreadable, that accumulated vocabulary is inert. You cannot search it, so you cannot reuse it, so in practice you rebuild it from memory each time. Recovering the prompts makes years of your own experimentation available again as something you can query.

This is what we mean when we talk about memory as infrastructure. The knowledge was already yours. It just wasn’t findable.

What to Do With a Recovered Archive

A few things become possible once old prompts are readable:

1. Search your own back catalogue by phrasing. Find the images where a particular construction worked, rather than scrolling for them.
2. Notice your own patterns. Recurring vocabulary across a few thousand prompts is a fairly honest portrait of how you work.
3. Import original downloads, untouched. An original download carries the prompt twice. Renaming a file loses the filename copy; re-saving or converting it can lose the embedded copy. An untouched download keeps both.

That last point is the practical one. If you are about to import a Midjourney archive, import the files exactly as they were downloaded.

Key Takeaways

• Midjourney writes your prompt twice: inside the image’s Description field and into the download filename.
• The embedded copy is complete but is stripped by re-saving, conversion, and some upload paths.
• The filename copy survives those, but truncates long prompts at download time and drops your parameters.
• Numonic now surfaces embedded prompts cleanly and falls back to filenames, for images already in your library.
• Keep original downloads untouched — renaming loses one copy, re-saving can lose the other.