Compression memory
Parts of an image saved at a different quality no longer match the rest of the frame. Error Level Analysis renders that disagreement as a heatmap you can read.
Forensic analysis
See what an edit leaves behind.
A manipulation leaves traces in the statistics of an image long after the visual evidence is gone. Twelve instruments read those traces — compression, noise, gradients, colour, repetition — and turn them into result images you can zoom, compare and export.
Twelve instruments · six recipes · eleven analysis panels
Reads JPEG, PNG, TIFF, WebP, GIF and BMP
Entirely offline — findings never leave the machine
Why the statistics, not the eye
Retouching is designed to survive the naked eye, and it usually does. What it finds harder to survive is arithmetic — every edit re-encodes, re-smooths or re-colours its region, and that leaves measurable footprints against its neighbours.
Noise floors
Sensor noise is remarkably even across a frame. Inserted content rarely matches the surroundings, and the difference is often visible in the noise alone.
Colour statistics
Splices arrive with someone else's colour history. Principal component analysis looks for colour structure a single photograph is unlikely to have.
An instrument, not an oracle
Every instrument shows you a map, not a sentence. You take the map to the next instrument, and the report is where the pieces come together.
The twelve instruments
Each instrument runs on the loaded image and produces a result you can zoom, compare and export — alongside the statistics behind what it saw.
| Instrument | What it reveals |
|---|---|
| Error Level Analysis (ELA) | Regions saved at a different compression quality — the classic signature of a spliced or retouched area. |
| Noise analysis | How the sensor noise is distributed. Inserted content rarely matches the surrounding noise floor. |
| Noise consistency | Whether the noise floor is uniform across the frame, which exposes compositing. |
| Noise residual | The signal left behind after denoising — useful for seeing structure hidden beneath texture. |
| SNR map | A per-region signal-to-noise view for locating locally anomalous areas. |
| Noise rainbow | Noise rendered as false colour, which makes otherwise invisible patterns visible at a glance. |
| Level sweep | Sweeps tonal levels to expose faint edges and boundaries. |
| Luminance gradient | Gradient and edge structure with selectable operators — good for spotting lighting that does not agree across a composite. |
| PCA | Principal component analysis over the colour channels, in several modes, for colour-level anomalies. |
| Block similarity | A block-level engine for copy-move detection — repeated regions within the image, found with four matching algorithms and clustered into results you can inspect. |
| Watermark detection | A frequency-domain scan for embedded watermark carriers — spectral evidence, not a search for logos or text. |
| Geo-tag audit | Finds location data recorded in the file, grades the quality of the fix, and writes map references into the report. |
Enhancement: make the faint visible
Analysis often begins with something barely there. Eight enhancement modes sit alongside the instruments, applied to a working copy — the loaded file itself is never rewritten by analysis.
- CLAHE — adaptive local contrast that lifts detail out of shadows and highlights without flattening the frame.
- Gamma correction — brighten or deepen the midtones without moving the endpoints.
- Levels — set the black point, white point and midtone for a deliberate tonal stretch.
- Unsharp mask — the classic edge lift, with amount, radius and threshold under your control.
- Local contrast — a gentler mid-frequency lift for reading texture across a region.
- Auto contrast — a one-click tonal stretch that finds the useful black and white points for you.
- Auto contrast (per channel) — the same stretch applied to each colour channel, for images with a colour cast.
- Histogram equalisation — redistributes the tonal range across the whole frame.
A lens, not a fix
None of this edits your photograph for good. The modes exist to make a faint structure legible, and they act on the working copy — the source file keeps its pixels exactly as they were.
Six recipes for six questions
Rather than configuring eleven panels by hand, a recipe enables the subset that answers your question — and collects the results into one summary.
| Recipe | The question it answers |
|---|---|
| Quick Triage | Is there anything here worth a second look? |
| Authenticity Check | Do the signals agree with each other — or contradict? |
| Provenance & Origin | Where has this file been, and what does it claim about where it came from? |
| Steganalysis | Are there signs in the image statistics of something hidden or modified? |
| AI Forensics | What do the synthetic-origin signals have to say? |
| Deep Dive | Everything at once, for the file that matters most. |
One score, banded honestly
The analysis resolves into a weighted authenticity score. It starts at 100 and subtracts for what it finds — high, medium and low-risk inconsistencies, hidden payloads, AI indicators, unverifiable provenance — with the breakdown shown so you can see which signal moved the number.
| Band | Range | What it means |
|---|---|---|
| Likely Authentic | 80 and above | Nothing the scan looked for spoke up. An absence of findings — not a guarantee. |
| Uncertain | 50 to 79 | Some signal fired, or the image is hard to read. Judgement withheld. |
| Suspicious | Below 50 | Findings stack up, or one is severe. Go to the instruments rather than the number. |
A summary, not a verdict on anyone
The score measures what the app could check on your machine. Read it as a pointer to where to look next — the contributing signals and their confidences are the actual evidence.
Findings you can hand over
A forensic result that cannot leave the app is not much use. The session exports as an HTML or PDF report, or as structured JSON — with the result imagery, verdicts and file digests attached.
| Output | Contents |
|---|---|
| Forensic report — HTML or PDF | Findings, verdicts and the result imagery, laid out as a document. |
| JSON report | The same content in structured form, for your own tools, records or review. |
| File digests | SHA-256 and SHA-512 integrity hashes, computed over the file and carried alongside the findings. |
| Result images | Analysis output saved as PNG, ready for use elsewhere. |
Reports come from the same engine the workflow system uses, so one produced by hand and one produced by a batch run contain the same information.
What a finding is — and is not
Forensics here means offline statistical measures, honestly labelled. The app is careful about the difference between a measurement and a conclusion; so is this page.
- Every instrument is a hand-written statistical measure — there is no cloud inference and no model file anywhere in the product.
- Checks of this kind produce false positives and false negatives in both directions.
- Metadata can be forged, stripped, or added innocently by an editor.
- A provenance manifest proves what the manifest says — not that a photograph depicts reality.
- A verdict that could not complete is reported as incomplete — never quietly treated as clean.
- An analytical finding is not a legal determination; admissibility is a question for your jurisdiction and your lawyer.
Use the signals, not the label
Before you repeat a verdict anywhere it matters, read which instruments contributed and what they showed. The three-tier label starts the conversation; the signals are what you stand on.
Frequently asked questions
Can StegoPixel prove that an image is fake?
No, and be wary of any offline tool that says it can. These are hand-written statistical heuristics, and every check of this kind produces both false positives and false negatives — the results are built to be read as evidence, with the contributing signals shown, not as proof.
Can I use the forensic reports as evidence?
The reports are built to be presentable: the HTML and PDF carry the result imagery, the verdicts and the file's SHA-256 and SHA-512 digests, which is a reasonable basis for internal review or a professional write-up. Whether they are admissible in a legal proceeding depends on your jurisdiction's rules on tool validation and expert testimony — a question for a lawyer, not a marketing page.
Does forensic analysis modify my image?
No. Analysis reads the file; enhancement and the instruments work on a working copy, and exported result images are new files. Your original is left exactly as it was.
Does analysing an image need an internet connection?
No. Everything runs locally — the app ships without an HTTP client or telemetry, so a file under analysis never leaves the machine. The only outbound links the product can produce are map references written into a geo-tag report, and those are contacted only if you click them.