Adding grain can change an image's appearance without making an AI detector accept it as a photograph. In Deprint's pilot, camera-style processing did not establish a reliable pass. Some low scores were misleading because the detector already missed the untreated originals.
If you like a grain effect, use it for the look. Treat a claim about detection as a separate question that needs its own evidence.
Separate texture from the final export
Illustrative pipeline. Isolate each step before attributing a result to grain.
Baseline
What this means
This checks whether the export itself changes the result.
Texture
What this means
Compare this with the control at the same dimensions and format.
Final delivery
What this means
A changed score now mixes texture, resizing and compression unless those factors have their own controls.
Why a camera look is not proof of camera capture
Grain, color response, blur and sharpening affect appearance. A visual AI detector may respond to other patterns, and a watermark verifier looks for a particular embedded signal. A convincing photographic style is therefore not a substitute for either check.
The pilot compared nine fixed treatments, including ordinary JPEG export, RGB grain, optical effects, simulated sensor noise and combined camera processing. It produced 207 variants from 23 files across 17 scene groups. Many files were related versions of the same scene.
What happened in the camera pilot
| Images and treatment | UniversalFakeDetect flagged as AI | NPR flagged as AI |
|---|---|---|
| 13 AI-labelled scenes, untreated | 2 of 13 | 13 of 13 |
| Same scenes, combined camera treatment | 0 of 13 | 12 of 13 |
| Same scenes, camera treatment plus JPEG | 2 of 13 | 12 of 13 |
| Four photographic controls, untreated | 0 of 4 | 2 of 4 |
| Same photographic controls, camera treatment | 0 of 4 | 2 of 4 |
UniversalFakeDetect's zero after camera treatment looks encouraging until you see that it missed 11 of the 13 originals. NPR detected more originals but also flagged two photographs. These are small exploratory counts, not representative accuracy estimates.
Sightengine separately checked 13 AI exports across two scenes, including originals and several processing variants. All returned 99% GenAI. A photographic control returned 0% GenAI.
The visual cost can be real
In two initial fixtures, adding virtual-camera processing after DeSynth reduced SSIM from 0.956 to 0.875 for the portrait and from 0.948 to 0.891 for the office scene. That means the measured images moved further from their originals; it does not quantify how every viewer would judge them.
A later, separate portrait experiment illustrates visible camera-stage changes:
Before means the restored intermediate here, not the original. The later camera stage changed color and texture. Both are native-size exports; the final resized JPEG is a separate file.
Select an image to inspect the full-size file. These are the archived files; your browser scales their display. Example from AI-Image-Detection-Bypass, with Deprint processing. Attribution and file hashes. No endorsement is implied.
This later example is not one of the pilot's 13-scene rows. Its Hive score fell after the camera stage, while the same full chain remained positive on desk and bakery scenes. Camera processing can interact with other operations; neither the pilot nor this selected example justifies an always-works or never-works claim.
Should you add grain to your image?
Choose grain if it improves the style you want. Check whether it obscures fine lettering, product textures or skin detail. Keep a clean original so the aesthetic choice remains reversible.
If your aim is to change a detector result, test the exact exported image and inspect its quality. A camera brand written into EXIF is not evidence that the image was captured by that camera, and these experiments did not fabricate camera identity.
Read why detectors can disagree before treating a single low score as confirmation.
Grain is an aesthetic choice before it is a detection claim
People associate some textures with film, high-ISO photography or a particular editing style. Adding such texture can change how a picture feels. That is a real creative effect, but it does not create a history of physical camera capture.
Nor does “camera-like” identify one operation. A treatment may combine noise, blur, sharpening, color changes and encoding. If the result changes, the whole bundle is responsible until separate controls isolate the parts.
A simple test would compare an unchanged export, grain only, other camera effects without grain, and the combined treatment. Keep dimensions and encoding consistent at that stage. Then add the final delivery export as a separate comparison.
Why low baseline detection changes the story
Suppose a detector misses most originals before any treatment. A near-zero count after treatment can sound impressive while adding little evidence of an effect. The relevant comparison is paired: what changed on the same scenes, under the same procedure?
In the pilot, UFD flagged only two of the thirteen untreated AI-labelled scenes. Its zero after the combined camera treatment should be read against that low baseline. NPR supplied a different view: it flagged all thirteen originals and twelve after treatment, while also flagging two of four photographic controls.
| Question | Needed comparison |
|---|---|
| Did the treatment lower a score? | Paired original and output scores |
| Did it change the label? | Paired decisions at the same threshold |
| Was the detector useful on this sample? | Original generated images and camera-photo controls |
| Did grain specifically cause the change? | Grain-only and matched non-grain controls |
| Does the effect generalize? | Fresh independent scenes with a frozen method |
No single row answers all the others. That is why the small pilot counts are reported directly rather than converted into a broad success claim.
Keep the visual tradeoff visible
Texture can conceal fine detail as well as add character. Inspect small labels, flat product colors and skin regions. A treatment that looks pleasing on a mood-board portrait may be unsuitable for a catalog image or a technical diagram.
The archived before/after example is a later portrait stage, clearly separate from the pilot. It helps readers inspect the appearance of that treatment without implying that every pilot output looked the same. The actual files are available at full size.
If you prefer the grain effect, that can be a valid reason to keep it. If the purpose is detection, measure the exact delivered file. A favorable visual judgment and a favorable detector result are separate outcomes, and either can occur without the other.
A future study should report both the detector changes and the proportion of outputs that still meet predefined quality needs. Adding more variants of the same few scenes would not, by itself, answer the generalization question.
How we tested, and what remains unknown
The pilot used fixed analytical settings, published detector preprocessing and original-file controls. The 13 AI scene labels came from upstream material and were not all independently verified. The four photo controls are too few to estimate a general false-positive rate. The tests were exploratory and did not include a blinded human realism study.
Observations were collected October 7–9, 2026. Methods and downloadable counts distinguish the pilot from later camera and portrait experiments.

