Skip to content
SynthIDX

Share a Flux Image Without the Workflow Data

Flux Metadata Remover

Strips the PNG metadata, prompts and workflow JSON in Flux exports, including ComfyUI pipelines. No visible watermark is added by Flux itself.

What a Flux export carries besides the picture

A FLUX image reaches your drive by two paths, and they leave different records behind. On the hosted side, Black Forest Labs signs generations made through the BFL API, the Developer Playground and BFL hosted integrations. On the local side, most people run the open weights through ComfyUI, and ComfyUI writes its own record into the file it saves. The two paths come from different teams and different documentation, and a single file can carry both if it passed through one environment and then the other. Nothing in the Flux documentation describes a visible mark on the output, which is why the picture itself usually arrives unmarked.

BFL attaches a C2PA Content Credential to every image its own pipeline returns. The credential is a signed manifest, embedded as a JUMBF block in JPEG and PNG output and as XMP in containers that cannot hold JUMBF natively. The BFL help centre article at https://help.bfl.ai/articles/3705384267-c2pa-in-flux-generated-images lists what it records: an assertion that the media is AI generated, the claim generator name with the signing software version, the creation event with a UTC timestamp, the FLUX model name and version, and a hash that binds the manifest to the pixels. The same article says the credential does not record text prompts, API keys, account identifiers, IP addresses or request metadata.

Self-hosted runs sit outside that system. The article states that outputs from self-hosted or on-premise deployments of the open weights are outside BFL signing, and that BFL does not attach or validate credentials for third party inference environments. A ComfyUI render made with FLUX.1 [dev] or [schnell] therefore arrives with no BFL credential, and the record it does carry comes from ComfyUI. That record is plain JSON inside PNG text chunks: a prompt field and a workflow field, both documented at https://docs.comfy.org/development/api-development/workflow-metadata.

The model family explains why both cases exist. FLUX.1 [pro] is served through the API only, [dev] ships as open weights under a non-commercial license, and [schnell] is Apache 2.0 and runs in a few steps, as the ComfyUI Flux tutorial describes at https://docs.comfy.org/tutorials/flux/flux-1-text-to-image. FLUX.2 adds the [klein], [pro], [flex] and [max] line covered in the official documentation at https://docs.bfl.ai/. The practical point: one model family can produce a file with a signed credential, a file with ComfyUI JSON, or a file with neither, depending on where it ran.

What the workflow JSON and the prompt JSON each record

The workflow field is the canvas you built. ComfyUI describes it as the complete graph, nodes, links and layout included, and the API format guide at https://docs.comfy.org/development/api-development/workflow-api-format lists what the save format keeps that an API export drops: node positions, node sizes, colors and groups. For a Flux pipeline the widget values in that graph name the checkpoint file, the two text encoders, the VAE, the sampler and scheduler, the guidance and step settings, and the seed. A reader who opens the text chunk sees the whole arrangement as readable JSON.

The prompt field is the execution form. ComfyUI calls it the API prompt used to run the workflow, and it holds numeric node ids, class types and an inputs dictionary while it leaves out muted or disconnected nodes. That is why the two records can differ in both shape and size: the prompt is compact and repetitive, and the workflow repeats the layout of every group. A saved image may carry one field without the other, depending on whether the frontend populated the extra_pnginfo dictionary before the save node ran.

Neither field bundles the things it names. The official metadata page puts this in a limitations list: the workflow JSON does not bundle the model files, input assets or custom node packages it depends on, and moving or renaming one of those can stop a recovered graph from running. What leaks is the recipe and the file names, not the multi gigabyte checkpoint the ComfyUI tutorial quotes for the original FLUX.1 weights. Sending a cleaned image does not hand anyone a working pipeline, and it does not hand over the model either.

Size shows the same split. A save-format workflow carries layout for every node and group, while the API export strips positions, colors and node sizes to stay small for queue submissions. When a Flux image comes back from a ComfyUI session, the embedded workflow is often the larger of the two records, and a few dozen nodes can push the file well past what its pixel dimensions suggest.

How to keep Flux workflow data out of an export

  • Turn metadata saving off inside ComfyUI

    Start ComfyUI with the --disable-metadata flag, or in ComfyUI Desktop enable the Disable saving prompt metadata in files option under the Server-Config panel. The official documentation notes that files made with this option on cannot be used to recover a workflow from embedded metadata. That is the trade: nothing is written, and dragging the image back onto the canvas no longer rebuilds the graph.

  • Keep the graph in a separate JSON file

    When you still want a record of the work, export the workflow as its own .json through File then Save, or through Export Workflow (API) for automation. The API guide makes the difference clear: the save format keeps layout and colors, the API format drops them for queue submission. A standalone file stays under your control and does not travel with every copy of the image.

  • Decide between hosted and local before you generate

    The same prompt yields a differently recorded file on each path. A BFL API or Playground image carries a signed credential, and the help centre says BFL does not offer its signing certificates to third parties. A ComfyUI install with the open weights carries none of that, and the documentation leaves content marking duties to the deployer. Choose the path that matches what you need to prove later.

  • Strip the exact copy you plan to share

    Once an image leaves your machine, the copies on other servers stop being yours to edit. Run the strip on the file you are about to send, and send that file. If you resize, crop or convert it afterwards, check the result again, because the editing app can write a record of its own into the new export.

How it works

How to strip metadata from a Flux PNG

  1. Step 1: Upload the Flux export

    Drop a PNG, JPEG or WebP file into the panel below, or click to browse. The file moves from your browser to processing storage and is not published anywhere. The copy on your disk stays untouched until a job runs, so you can compare both files afterwards if you want to.

    Tip

    Start from the file the generator or ComfyUI wrote. A copy saved out of a chat thread may have lost the record you wanted to inspect.

  2. Step 2: Run the strip

    The panel starts the image metadata pass with removal already switched on, so there is nothing to configure. The job clears the generator records it recognizes: a prompt or workflow text chunk, an extra_pnginfo entry, a parameters style block, and any C2PA manifest whether it arrived as a JUMBF block or as XMP. On JPEG the cleaner walks the segment table and drops the metadata segments without touching the coded image, a PNG above 8 bits per channel keeps its compressed image data as it was, an 8 bit PNG is resaved by an encoder and does not change the pixels, and a WebP is re-encoded when the cleaned file is written.

    Tip

    Credits are spent only when a strip job actually runs.

  3. Step 3: Download the cleaned file

    When the job ends, take the cleaned image from the download link and open it once more before it travels. The frame should match what you uploaded, and the color profile should keep colors where they were. Upload that copy straight to wherever you plan to post it. A signed-out visitor is asked to sign in first, and the panel returns once that is done.

    Tip

    Run the strip again after any edit. The app that saved the new file may have written its own record into it.

Limitations and what the strip leaves in the file

A soft binding may survive in the pixels

BFL documents an optional soft binding, described as an invisible watermark that links an image to its credentials even after the container is stripped or re-encoded. The strip edits metadata sections around the frame, and it cannot reach a signal woven into the pixels. There is no guarantee the cleaned file is unlinkable, and the page makes no such promise.

The credential is gone, and absence is visible

After the strip, a C2PA verifier reports that no credential was found. That is the intended result for a file you own, and a recipient can observe the same absence. Decide whether that is the outcome you want before you run the pass.

A visible mark is not metadata

Flux itself adds no visible mark, but a third party integration, a hosted hub or a finishing step can paint a logo, a signature or a border into the pixels. Nothing in a metadata pass changes those pixels, so a painted mark remains exactly where it was drawn.

Standard metadata stays in place

The default pass targets records that name an AI generator. It does not remove genuine camera fields, capture timestamps, GPS coordinates or the embedded color profile, and the profile is kept so colors do not shift. To clear the personal layer as well, use the Remove Metadata page, which runs a separate privacy pass on the same image.

Some Flux files have nothing to strip

A ComfyUI render of an open weight model is outside BFL signing, and a session started with metadata saving off writes no chunks. When that happens the job reports nothing to remove. A clean report is a valid result, and it says nothing about how the file was handled before it reached you.

Copies already shared cannot be recalled

The pass works on the file you upload. It cannot edit an image you sent last week, a download someone else saved, or a copy a platform keeps on its side. Cleaning an export before it travels is the step within reach, and cleaning afterwards only helps the copies you still hold.

A platform may rewrite the file after upload

Many platforms re-encode an upload into their own delivery format. That rewrite may drop some fields and add others. The record you can control is the one in the file you send, and what a platform serves later belongs to the platform.

The decision to remove a record belongs to its owner

Prompts, graphs and credentials exist to record how an image came to be. Clearing them from your own output is ordinary editing. Clearing them from another person’s image so that it reads as yours misstates where it came from, and it may break their terms or local law. Work on files you made or were given permission to change.

Identity and official documents are not supported

This tool is not for passports, identity cards, driver licenses, visas, certificates, statements, or other official documents, whether the document is real, generated by AI, or put together from a template. Editing one of these can be document fraud, and we close accounts that try. Use the tool on other images you generated or have the right to edit.

Where to go next

The Stable Diffusion Metadata Remover page covers the sibling apps that write PNG info chunks, including the parameters block that Automatic1111 and Forge produce, for anyone who works with both model families. The AI Metadata Remover page takes the wider view of generator records, for images that did not come out of a Flux pipeline and for the IPTC and XMP fields a hosted service may add. The Remove Metadata page takes the other layer, the location, device and capture time fields this page leaves untouched.

The facts on this page come from official sources. The BFL help centre describes where Content Credentials are attached and what they record, and it also covers self-hosted and third party deployments, at https://help.bfl.ai/articles/3705384267-c2pa-in-flux-generated-images. BFL model and API documentation is at https://docs.bfl.ai/. The ComfyUI metadata page describes the prompt and workflow fields, the supported containers and the --disable-metadata option at https://docs.comfy.org/development/api-development/workflow-metadata, and the API format page at https://docs.comfy.org/development/api-development/workflow-api-format explains the two export shapes. The C2PA standard itself is published at https://c2pa.org/.

Sign in to run. Free credits included.

FAQ

Frequently asked questions

Short answers about the prompts, workflow graphs and signed credentials in a Flux export, and what a metadata strip can change.

Explore more tools

Stable Diffusion Metadata Remover

The same job for the Automatic1111 and Forge apps, which pack the prompt and sampler settings into one parameters block.

Open tool

AI Metadata Remover

The general entry for generator records from any service, for images whose origin is not a Flux pipeline.

Open tool

Remove Metadata

The privacy pass for location, device and timestamp fields, for the layer this page keeps by default.

Open tool

Image Metadata Viewer

See the workflow and prompt fields inside a Flux PNG before the strip.

Open tool

Grok Metadata Remover

A second platform where the visible watermark is documented and the metadata is what this tool clears.

Open tool

All tools

Every remover and detector on one page, grouped by what each one cleans or checks.

Open tool

Ship a Flux image without the recipe attached

One metadata pass clears the ComfyUI prompt and workflow chunks and any C2PA credential, while PNG and JPEG keep their decoded picture unchanged.