Removing Watermarks From Midjourney, Firefly, and Stable Diffusion Images

Not every AI image carries the same kind of watermark. A picture from Midjourney, one from Adobe Firefly, and one from Stable Diffusion can each look identical on screen while hiding completely different marks in the file. Treating them the same way is how people end up publishing an image they thought was clean. This guide walks through the three generators one at a time, because the right fix for a corner logo is not the right fix for embedded provenance metadata.

#Three layers, three different fixes

Every generator can mark an image on up to three layers, and each layer needs its own approach. Getting this framing right first will save you from guessing. Our complete guide to AI watermark removers covers the full picture, but here is the short version.

1. Visible marks: logos, sparkles, signatures, or text you can see. Fixed by cropping, inpainting, or regenerating the affected area.
2. Metadata and provenance: EXIF fields, PNG text chunks, generation parameters, and C2PA Content Credentials stored in the file container. Removed by re-encoding or stripping metadata.
3. Embedded invisible signals: pixel-level watermarks such as Google SynthID that survive cropping and compression. Only disrupted by diffusion or SDXL regeneration, and never guaranteed.

#Midjourney: mostly your own overlays

Midjourney does not stamp a forced logo onto standard output by default, so most visible marks on a Midjourney image are ones the creator added: a signature, a studio name, a caption, or a border applied after download. Those are ordinary pixel edits. If the mark sits over a flat background it usually reconstructs cleanly; if it overlaps a face, fine text, or a busy edge, inspect the repair closely at full size.

Midjourney images also carry metadata worth checking. Depending on how a file was exported and where it passed through, it may include software tags or platform information in the container. A quick pass through a watermark remover that re-encodes the file clears most of that, but treat metadata and any embedded signal as separate questions from the visible overlay.

#Adobe Firefly: Content Credentials by design

Firefly is the outlier here. Adobe is a founding member of the Content Authenticity Initiative, and Firefly output is designed to embed C2PA Content Credentials that record how the image was made and edited. This is provenance metadata, not a visible logo, so you will not see it on the image itself. You can inspect it at contentcredentials.org/verify by dropping the file in.

Because Content Credentials live in the file container, re-encoding or explicitly stripping metadata removes the embedded manifest. Keep in mind that provenance systems can also reference records stored elsewhere, so a clean file does not prove every trace of origin is gone. If you are editing a Firefly image, decide honestly whether disclosure is required for how you plan to use it.

#Stable Diffusion and SDXL: metadata plus whatever the pipeline added

Stable Diffusion and SDXL are the most variable case because so much depends on the specific pipeline. Images generated through common tools frequently carry generation parameters in the file: a Software tag, the prompt, sampler settings, seeds, and model hashes written into EXIF or PNG text chunks. Anyone who opens the file in the right viewer can read exactly how it was made.

On top of that, some open pipelines add an invisible watermark of their own, and a growing number of setups attach C2PA data. So a single SDXL image can touch all three layers at once: no visible mark, a full block of parameter metadata, and a faint embedded signal. Stripping the metadata is straightforward through re-encoding; disrupting an embedded invisible watermark requires regeneration and is never guaranteed against every proprietary detector.

#How to tell which layer you are dealing with

Before you edit, spend a minute identifying what the image actually carries. This is faster than it sounds and stops you from over-editing an image that only needed a metadata strip, or under-editing one that still holds an embedded signal.

• Look at the image at full size. Any logo, signature, sparkle, or caption is a visible-layer problem.
• Drop the file at contentcredentials.org/verify to see whether it carries C2PA Content Credentials. Expect this for Firefly and some SDXL pipelines.
• Inspect EXIF and PNG text with a metadata tool to reveal generation parameters, common in Stable Diffusion and SDXL exports.
• For an embedded signal, use Google’s SynthID Detector where applicable; remember no consumer tool can certify that every proprietary detector is defeated.

Once you know the layers, the fix follows. A visible signature needs inpainting; parameter metadata needs a re-encode; an embedded watermark needs regeneration with the understanding that results vary. If you want more on the invisible layer specifically, see what SynthID is and how removal works in practice. And if you are unsure whether your intended edit is appropriate, our note on the legal and ethical side is worth a read first.