Midjourney Prompt Metadata in 2026: What's Embedded, What Leaks, and What Breaks in Batch Exports

Download any image from midjourney.com today—single or batch—and it arrives with your full prompt, generation parameters, Job ID, author, and an IPTC digital source tag, all written into the PNG metadata. That is a genuine improvement over 2024, when downloaded files were blank slates. Midjourney now embeds more provenance data than most users realise.

But “more than nothing” is not the same as “enough.” The metadata is packed into a single Description field as a text string. There are no separate, machine-readable fields for seed, model version, or parameters. There is no variation or upscale lineage. And there are no C2PA Content Credentials. This article documents exactly what Midjourney embeds as of March 2026, what it means for privacy, what your older downloads are missing, and where the gaps remain.

What Changed in Late 2025

Prior to approximately October 2025, every image downloaded from Midjourney—whether through Discord or the web app—contained no generation metadata at all. The PNG carried standard EXIF fields (dimensions, colour space, creation date) and nothing else. Your prompt, seed, model version, aspect ratio, and style reference existed only in Discord messages or the midjourney.com generation history.

This changed when Midjourney began embedding generation parameters into downloaded PNGs. The improvement was rolled out without formal announcement, first noticed by community members inspecting EXIF data in late Q4 2025. For users downloading one image at a time from the web app, the metadata is now meaningfully complete.

The catch: the implementation is inconsistent across download methods. That inconsistency is the subject of this article.

What's Actually Embedded (Verified March 2026)

We inspected PNGs downloaded from midjourney.com in March 2026—both a single download and an image extracted from a batch ZIP. Both contained identical metadata fields. Here is exactly what Midjourney embeds:

• Description — A single text field containing the full prompt, all parameters (--ar, --stylize, --chaos, --weird, --v), and the Job ID. If you used an image reference URL, that appears too. Everything is packed into one string.
• Digital Image GUID — The Job ID stored as a proper UUID field, separate from the Description text. This enables programmatic matching against your midjourney.com generation history.
• Author — Your Midjourney username (e.g., dr.jb). This travels with every file you share.
• Creation Time — The timestamp of the original generation, not the download time.
• Digital Source Type — The IPTC standard value trainedAlgorithmicMedia, which marks the image as AI-generated using an industry standard vocabulary. This is notable—Midjourney does use at least one recognised provenance standard.
• XMP Toolkit — Identifies the XMP implementation (XMP Core 4.4.0-Exiv2).

This is more than many users expect, and it is a substantial improvement over 2024. A downloaded PNG from midjourney.com is genuinely useful as a self-documenting creative asset. You can inspect the file, recover the full prompt, and identify the exact generation job.

What's Not Separately Structured

The good news is that the data is there. The nuance is how it is stored. Midjourney packs almost everything into the Description field as a single text string. There are no separate, machine-readable fields for individual parameters:

• No separate seed field — If you used --seed, the value appears in the Description text. If you did not specify a seed, it is not recorded at all. There is no dedicated EXIF field for it.
• No separate parameter fields — Aspect ratio, stylize, chaos, weird, and model version are all part of the Description string. Extracting --stylize 700 requires text parsing, not metadata lookup.
• No style reference identifier — If you used --sref, the reference URL or code appears in the Description, but there is no structured field linking to the style.

This matters for automated workflows. A DAM system, script, or search tool that wants to filter your library by “all images generated with --v 7 and --stylize above 500” must parse a free-text string rather than reading structured fields. It works, but it is fragile—any change to Midjourney's Description format would break downstream parsing.

Single vs Batch: The Same Metadata (as of March 2026)

Earlier community reports suggested that batch ZIP downloads stripped metadata that single downloads preserved. Our own testing in March 2026 found identical metadata fields in both single and batch-downloaded images. The Description, Digital Image GUID, Author, Creation Time, and Digital Source Type were all present in files extracted from a batch ZIP.

It is possible that earlier batch exports behaved differently, or that specific edge cases (very large batches, older generations, video files) still produce inconsistent results. But as of this writing, the single-vs-batch metadata gap appears to have been resolved for PNG downloads. If you experienced metadata stripping in batch exports previously, it is worth re-testing with a fresh download.

Important caveat: We tested a recent batch download in March 2026. Images downloaded in earlier months or years may not have had this metadata embedded at the time of download. The metadata improvement was rolled out gradually, and re-downloading the same image today may yield richer metadata than the file you saved six months ago.

Metadata by Download Method: A Comparison

Two things stand out. First, single and batch downloads now contain identical metadata—the earlier inconsistency appears resolved. Second, while Midjourney does use the IPTC Digital Source Type standard to mark images as AI-generated, it does not embed C2PA Content Credentials or preserve variation/upscale lineage. The IPTC tag is a genuine step toward provenance, but it is not cryptographically signed and it does not meet the full requirements of the EU AI Act.

The Privacy Problem: Your Prompts Are Showing

There is a flip side to embedded metadata. For every user who wants their prompts preserved, another wants them hidden. When you share a Midjourney image with a client, post it on social media, or include it in a portfolio, the embedded prompt travels with the file.

This creates concrete problems:

• Client confidentiality — A prompt referencing a client's brand name, product details, or creative brief is now readable by anyone who inspects the file
• Creative IP — Carefully crafted prompt techniques are embedded in the output, visible to anyone who knows where to look
• AI disclosure — In contexts where you want to use AI imagery without highlighting the method, embedded prompts are an involuntary disclosure
• Social media scraping — Uploaded images may retain EXIF data depending on the platform, making prompts discoverable at scale

Midjourney does not offer a metadata-stripping option at download time. If you want to share an image without its prompt, you must strip the EXIF yourself—or use a tool that does it for you with configurable privacy presets. This is one of the reasons a post-export management system matters: you need control over what metadata leaves your library and what stays internal.

Your Pre-2025 Archive: The Metadata Desert

If you have been using Midjourney since 2022 or 2023, you likely have hundreds or thousands of images downloaded before the metadata improvement. These files contain nothing—no prompt, no seed, no parameters, no job ID. They are visually rich but contextually blank.

Reconnecting these legacy files to their generation history requires one of two approaches:

1. Job ID matching — If the filename includes a Midjourney job ID (the UUID in the original filename), it can be matched against the midjourney.com generation history. This works well but depends on you having preserved original filenames.
2. Visual matching — For renamed files or screenshots, visual similarity search can identify the original generation in your MJ account, allowing metadata to be recovered and reattached.

Neither approach is available natively in Midjourney's web app. This is where a dedicated archive strategy becomes essential—especially for users whose legacy library represents months or years of creative work.

What Midjourney Metadata Is Not

To Midjourney's credit, the embedded metadata uses recognised standards. The Digital Source Type field carries the IPTC value trainedAlgorithmicMedia—an industry-standard machine-readable flag that this image was AI-generated. This is a genuine step toward disclosure. But it is important to understand what it is not.

• Not C2PA — Midjourney does not embed C2PA Content Credentials, the Coalition for Content Provenance and Authenticity standard adopted by Adobe, Microsoft, and others. C2PA provides cryptographically signed, tamper-evident provenance. MJ's IPTC/XMP metadata is unsigned and trivially editable—anyone with ExifTool can strip or alter it.
• Partial EU AI Act alignment — The EU AI Act (Article 50, enforcement begins August 2026) requires AI-generated content to be “marked in a machine-readable format and detectable as artificially generated or manipulated.” The IPTC Digital Source Type field is machine-readable and standards-based, which is a start. But because it is not cryptographically signed, it can be silently removed—so it likely falls short of full compliance on its own.
• Not lineage — Even complete metadata describes a single generation. It does not capture the variation–upscale–remix tree that led to the final image. Lineage requires a graph structure, not flat key-value pairs.

How to Check Your Own Files

You can verify the metadata state of your Midjourney library right now. Here is how to audit a sample:

1. Pick 10 images from different time periods and download methods (single, batch, Discord-saved, pre-2025)
2. Inspect EXIF using ExifTool (exiftool image.png), an online EXIF viewer, or any metadata inspection tool
3. Look for generation fields — Prompt text, seed, parameters, and job ID are stored in EXIF comment or description fields
4. Compare results — Recent downloads (single or batch) should have identical metadata. Anything from before late 2025 is likely blank. Discord-saved images may vary.

For a faster approach, use Numonic's MJ Export Metadata Inspector to upload a batch and see what's embedded versus missing—with privacy risk flags for files that contain prompts you may not want to share.

Bridging the Gap

Midjourney's metadata improvements are real and welcome. For users who download images one at a time from the web app, the experience has improved meaningfully since 2024. But for professional users managing thousands of images across batch exports, legacy archives, and multi-tool workflows, the metadata picture is fragmented.

The path forward involves three things: verifying what your current archive actually contains, establishing a reliable export workflow that preserves metadata, and adding a post-export management layer that handles what Midjourney cannot—privacy-aware sharing, legacy recovery, lineage tracking, and compliance-grade provenance.

Whether you use Numonic, ExifTool, or a manual inspection workflow, the first step is the same: know what your files actually contain. Current downloads have consistent metadata, but your back catalogue almost certainly does not—and that gap is the real challenge.