INITIALISATION DU SYSTÈME...
Automatic1111 and ComfyUI write the entire generation context into the PNG tEXt chunks: prompt, negative prompt, sampler, CFG, seed, model hash, LoRA stack, even ControlNet inputs. Anyone with a free PNG inspector reads it back word-for-word. If you ship Stable Diffusion outputs to a client or sell them on a stock site, you also ship your full workflow. GhostMeta wipes every tEXt and iTXt chunk via canvas re-encode — the visible pixels stay, the recipe disappears. Unlike cloud generators, a local Stable Diffusion install signs nothing — there is no JUMBF box, no APP11 C2PA segment, so the only leak surface is the freeform PNG text. But that text is plaintext, not hashed: the model hash alone resolves to the exact checkpoint on Civitai, and a marketplace reviewer can reconstruct your whole pipeline from one download. If you instead export to JPEG, Automatic1111 moves the same generation string into the EXIF UserComment field inside the APP1 segment, so it follows the file just as readily. GhostMeta decodes the bitmap to canvas and re-encodes it, so neither the PNG tEXt chunks nor the JPEG APP1 EXIF survive the round trip.
Stable Diffusion front-ends write the prompt, negative prompt, seed, sampler and checkpoint name into the PNG text chunks. Sharing the image means sharing the recipe. GhostMeta wipes those chunks along with the other layers, in the browser: clean the published copy, keep your original.
Common Stable Diffusion front-ends write generation parameters into the PNG text chunks. You will find the full prompt, the negative prompt, the seed, the step count, the sampler, the CFG scale and the name of the checkpoint or LoRA used.
This is plain text, readable by any editor able to open PNG metadata. Sharing the image means sharing the recipe: whoever picks it up can reproduce the render, and infer from the model list what your setup contains.
Running a model on your own machine protects you on the server side: nothing is sent anywhere during generation. It does not make the file discreet, because the local front-end itself writes the parameters into the PNG.
Cleaning removes those chunks along with the other layers. Check with the viewer first: on a PNG from a local interface, the parameters show up in plain text, which is the best argument for making a habit of cleaning before publishing.
Wiping the PNG chunks also wipes your working memory: the seed and the prompt vanish from the file, and with them the ability to reproduce the image later.
So clean the copy meant for publication, not your original. The sound habit is to keep the raw file in a local folder and to distribute cleaned copies only.
| Data | What it reveals | Removed |
|---|---|---|
| Prompt and negative prompt | Your exact wording | Yes |
| Seed | Allows the image to be reproduced | Yes |
| Checkpoint, VAE, LoRA | What your setup contains | Yes |
| Steps, sampler, CFG | Generation settings | Yes |
| Workflow graph (ComfyUI) | The whole processing chain | Yes |
ComfyUI writes the full graph as JSON in a 'workflow' tEXt chunk. GhostMeta drops it along with everything else when re-encoding.
Yes — the metadata format is the same PNG chunk structure across SD 1.5, SDXL, and SD3 outputs from A1111 and ComfyUI.
Not currently. GhostMeta is all-or-nothing on metadata to keep the privacy promise simple. Use a dedicated EXIF editor if you need partial preservation.
Automatic1111 writes the full generation string into a single PNG tEXt chunk keyed 'parameters' — one block holding prompt, negative prompt, seed, sampler, CFG and hashes. GhostMeta does not depend on that key: it enumerates every tEXt and iTXt chunk in the PNG stream, including compressed iTXt, and drops all of them on re-encode whatever they are named. So a renamed or non-standard key is stripped just the same.
The 10-character model hash and any LoRA hashes in the parameters chunk are public identifiers. Anyone can paste them into Civitai's search and land on the exact checkpoint and LoRA you used, exposing the style source you may have wanted to keep private. Stripping the text chunk removes the hash along with the seed, sampler and CFG.
No. When A1111 saves a JPEG it cannot use PNG text chunks, so it writes the same parameters string into the EXIF UserComment tag inside the JPEG APP1 segment. The data is just relocated, not removed. GhostMeta strips the APP1 EXIF block on re-encode, so the JPEG path is covered the same as PNG.
Partner : M.E Expert Serrurier