Deprint image-processing research: public methods and observations Published 9 October 2026. Observation window: 7–9 October 2026. Purpose and scope These articles explain broader questions using a small exploratory dataset. They do not generalize the results to all images. Source cutoff: 900e17467dcb05d417f20bdece597fecac0277ab; archive revision: ad5eddf3c7f4eb475b4a239bf3d61bb6c627f433. The internal archive indexes 27 experiment families and 31 reports. Later generalization work is excluded. Account details, browser screenshots and infrastructure records are not published here. Report names below identify frozen source records. Downloads in this directory observations.csv — selected measurements used by the articles. media-manifest.json — SHA-256 hashes and byte sizes of the displayed files. media-notes.txt and example-license.txt — source, attribution and MIT notice. portrait-original.jpg, portrait-restored.png, portrait-camera.png, portrait-final.jpg — exact archived image bytes; browser display scaling does not change those bytes. A. Watermark and visual-detector disagreement Source: camera-detector-pilot-2026-10-07.md, including 8 October follow-up. Google-origin desk and bakery scenes were generated and downloaded for testing. Metadata-free PNG controls preserved original decoded RGB pixels and triggered positive Gemini watermark observations. DeSynth outputs received negative watermark observations; Sightengine displayed 99% GenAI on both original JPEGs and both outputs. The PNG controls were not the Sightengine original files. Desk: 2400 x 1792, output SSIM 0.932. Bakery: 2528 x 1696, SSIM 0.897. Gemini referred to a PNG as JPG; internal upload conversion is unknown. Generated explanations about origin are excluded; only the reported check is retained. Two controlled observations do not establish a general removal rate. B. Selected portrait and fixed-chain transfer Sources: deai-chain-2026-10-08.md; restored-portrait-2026-10-09.md; production-gemini-2026-10-09.md. Public AI-attributed portrait source: AI-Image-Detection-Bypass at c1a80efadada67da16cdb7472b4681aaa7f20047. Deprint executed pinned SDXL redraw, a JPEG intermediate, detail restoration, camera processing and final export, without fabricated camera EXIF or the author's permissive detector parser. Broader portrait origin is not independently established. Native-stage Hive scores: redraw 92.4%; JPEG/restoration 42.2%; camera 7.1%. SSIM against original: 0.896663, 0.949589, 0.791954. These are cumulative stages, not a factorial ablation. Restoration includes the prior JPEG step. Final JPEG: Hive 2.9%. Its SSIM 0.819185 uses a resized reference and cannot be ranked directly against native SSIM. Original dimensions 768 x 1376; final 878 x 1573. A resize/JPEG-only original control stayed Hive 99.9%. Exact final SHA-256: b5a76e274abcc06ef6ebf4df0260b2c7822163c51f4a8bbc1949dcf13d57901d. It received Gemini 'not detected', Hive 2.9% negative classification and Sightengine 0% GenAI (separate face-manipulation indicator 1%). Positive original-pixel control SHA-256: df3967de13c6e6d09cc624f6c4552e4f62a7a066ba25308cbd41dfc89dce127b. The original JPEG's C2PA response alone was not used as a positive pixel-watermark control. 9 October timing: Hive 16:27 UTC, Sightengine 16:30 UTC, Gemini 17:11 UTC. Earlier receipts recorded that no joint watermark result yet existed; the later exact-file Gemini check supplied it. Qualified watermark observations came from Gemini's web Verify AI interface, not a successful standalone SynthID portal test or raw detector API. Internal preprocessing is unknown. Restored intermediate: Gemini negative, fresh Hive 42.2% negative classification, Sightengine not checked. Camera-stage Gemini not submitted after quota; diffusion-only Gemini unavailable. Missing checks are not negatives. The same fixed chain failed on desk/bakery: original controls Hive 99.9% each, final 90.9%/99.9%; native SSIM 0.809877/0.739094. Restored intermediates Hive 99.8%/99.9%. Selected/reused scenes cannot supply a representative success rate. C. Delivery-size controls Source: deai-size-control-2026-10-09.md. Originals and processed outputs were resized to approximately the portrait's delivered pixel area, preserving aspect ratio, with matching JPEG97 export. Desk (1360 x 1015): Sightengine 99% before and after. Bakery (1435 x 963): Hive 99.9% before and after. Bakery Sightengine checks were quota-unavailable. Different detectors cannot be pooled. Other resize settings were not exhaustively tested. D. Targeted local classifiers Source: bounded-detector-transfer-2026-10-07.md, including ensemble follow-up. Two DeSynth-processed scenes received bounded edits optimized first for NPR, then also SigLIP2/DINOv2. First study: 20 exports. Follow-up: eight exports. Both are variants of two scenes, not independent samples. Final PNGs were below both local thresholds; SigLIP/DINO 0.1500% desk, 0.0119% bakery. Hive stayed 99.9% for each. A fixed earliest-snapshot selection rule preceded Hive feedback. Both local classifiers influenced processing and are not independent judges. Final SSIM against DeSynth input: 0.98158/0.98634. Against original: 0.91559/0.88543. High similarity of the final edit is not full-pipeline preservation. Fresh negative Gemini reports did not establish independent visual-detector success. E. Analytical camera pilot Source: camera-detector-pilot-2026-10-07.md. Nine fixed arms: normalization, JPEG95, gamma roundtrip, grain, color-filter-array simulation, optics, sensor noise, combined camera and combined camera plus JPEG. 207 outputs from 23 files in 17 scene groups: 13 upstream AI-labelled scenes, four historical photos and six related intermediates. AI source labels not all independently verified. Settings frozen before scoring. UFD used its author's 224-pixel center crop/released head. NPR used its author's 256-square bilinear preprocessing/released diffusion-test checkpoint. Thresholds stayed 0.5. UFD/NPR flags: untreated AI 2/13 and 13/13; camera 0/13 and 12/13; camera/JPEG 2/13 and 12/13. Photos untreated: 0/4 and 2/4; camera-treated: 0/4 and 2/4. Not representative accuracy or false-positive rates. Separate Sightengine sample: 13 AI exports across two scenes all 99% GenAI; one photo control 0%. Do not confuse those files with the 13-scene open-detector panel. Initial DeSynth vs DeSynth-plus-camera SSIM: portrait 0.956/0.875; office 0.948/0.891. The later public portrait illustration is a separate experiment, explicitly labelled in the article. F. Stronger redraws Sources: desynth-strong-2026-10-09.md; qwen-sdxl-chain-2026-10-09.md; sdxl-strong-2026-10-09.md. Reused desk scene: stronger Qwen Hive 99.9%, SSIM 0.623759; Qwen then SDXL 99.8%, 0.610816; stronger SDXL 99.7%, 0.762468. Earlier fixed SDXL chain 90.9%, 0.809877. Different methods are not one isolated strength variable. Stronger staged SDXL small-fixture qualification failed strict pixel equality with the ordinary pipeline by at most one 8-bit value. A bounded numerical criterion was accepted afterward for research, not preregistered as exact parity. No universal conclusion follows. G. Selected JPEG metadata controls Source: deai-sample-check-2026-10-08.md. Original Hive 99.9%; APP11-stripped copy 99.9%. Author's processed sample 10.6%; APP1-stripped copy 10.6%. Compressed JPEG scan bytes unchanged within each pair; only selected segments removed. Author output differs from Deprint's later 2.9% output. This rules out the tested metadata-only explanation for that pair, not every metadata/detector interaction. Product Lossless delivery separately removed supported metadata with identical decoded pixels. This is not byte-identical file output or proof of pixel-watermark removal. No general compression-ratio benchmark is claimed. Interpretation Website scores are rounded and not calibrated origin probabilities. Services may change. Variants share scenes; errors, quotas and unsubmitted checks are missing evidence, never passes. Negative observations do not prove photographic origin, permanent watermark absence, preserved identity or future platform acceptance. No platform-labeling study, representative removal rate or broad commercial-tool ranking is claimed. Visual observations were not a blinded human study.