Thermostatic/community-forensics-low-quality-detector-2026-08

Community Forensics low-quality detector (August 2026)

This is a low-resolution-focused continuation of `Thermostatic/community-forensics-frontier-detector-2026-08`, which was independently fine-tuned from the MIT-licensed `OwensLab/commfor-model-384`. It is not an official OwensLab release.

The model is a fallible whole-image screening signal, not proof of authorship or provenance. It materially improves detection after ordinary downscaling, but it remains unreliable after severe multi-hop laundering and against small synthetic composites. The failed robustness gates are published rather than hidden.

Files

• —model.safetensors: canonical EMA inference weights.
• —community_forensics_low_quality_fp16.onnx: recommended browser/deployment artifact.
• —config.json and calibrator.json: preprocessing and frozen decision rule.
• —reports/final_report.json: standard calibration, recent-model diagnostics, robustness, and ONNX parity.
• —reports/promotion_report.json: complete baseline-versus-candidate low-quality gate results.
• —reports/redteam_report.json and .md: complete 33-condition development red-team evaluation.
• —metadata/training_summary.json: training curve and optimizer/run metadata.
• —configs/low_quality.yaml and configs/redteam.yaml: exact evaluation registries.
• —checkpoints/frontier_initial_best.pt: exact full optimizer/checkpoint payload used to initialize the continuation, SHA-256 e3259da45b0791ef9f43aa07ceff95ccd590c14c98cb3bcdc573bad23f06de1b.
• —checkpoints/low_quality_selected_last.pt: exact selected step-2,400 checkpoint, including optimizer, scheduler, EMA, RNG, and training history, SHA-256 fbe38867f4873882a04003e9f598b07735d183edfe5843d07ebbbc01d5ba166c.
• —reproduction/manifests/combined_manifest_inventory.csv.gz: path-free inventory of all 113,472 rows and their exact/perceptual hashes. Upstream image bytes are not redistributed where source terms do not permit it.
• —reproduction/manifests/combined_manifest_original.csv.gz: byte-exact original manifest, including the /data/ai_blink paths used by the resumable checkpoint. Decompression SHA-256 is af629187fca31ab2f4f7f4c81ea3a761c13ea7cb8d4ee8a9bb9e7bd8e1eb6f36.
• —reproduction/manifests/recent_holdout_inventory.csv.gz and openrouter_90_inventory.csv.gz: protected-cohort inventories.
• —reproduction/manifests/*_original.csv.gz and original_manifest_proof.json: deterministic compressed copies of the exact diagnostic and red-team manifests plus decompression hashes.

Exact training/evaluation source, tests, pinned environment records, portable commands, and manifest-rehydration tooling are maintained at `Sekinal/local-ai-image-signal-training`.

Preprocessing and decision rule

• —RGB input; resize short edge to 440 pixels; center-crop to 384 × 384.
• —Normalize with ImageNet mean [0.485, 0.456, 0.406] and standard deviation [0.229, 0.224, 0.225].
• —The model returns one raw logit x.
• —Calibrated score: sigmoid(0.6352260751077209 * x - 0.2643220522904507).
• —The frozen score threshold is 0.65, equivalent to raw logit boundary 1.390625.
• —The ONNX artifact has FP16 weights and FP32 input/output.

Dataset and training

The unchanged combined manifest contains 113,472 rows: 109,560 training images (41,313 real and 68,247 generated) and 3,912 calibration images (1,983 real and 1,929 generated). It spans five real-image training sources and 131 synthetic source buckets. The synthetic data includes both older generators and recent GPT Image, DALL-E 3, FLUX 1/2, Imagen 3/4, Seedream 3/4/5, Qwen Image, Hunyuan Image, Midjourney, Recraft, Ideogram, Janus, Sana, Z-Image, and related buckets.

The public frontier component is `Thermostatic/frontier-synthetic-images-2026`. It is not the full combined corpus; source-specific licenses and terms remain applicable. The combined manifest SHA-256 is af629187fca31ab2f4f7f4c81ea3a761c13ea7cb8d4ee8a9bb9e7bd8e1eb6f36.

The low-quality continuation initialized from the prior released detector and ran on one NVIDIA L40S:

The class-symmetric augmentation mixture applied identically to real and generated examples: 30% of samples used the prior web-laundering mixture, while 70% used tiny-source downscaling between 32 and 256 pixels with randomized interpolation, JPEG/WebP recompression, blur, noise, and unsharp filtering. The run reduced recorded training loss from 1.208 at step 1 to 0.268 at step 2,400. Selection and calibration used only the development/calibration role; the locked competition test was never opened.

Standard evaluation

All figures are development/calibration results on the 3,912-image group-disjoint calibration set.

The out-of-fold clean/web/hard macro balanced accuracy is 0.9264. These results are development evidence, not an untouched competition-test estimate.

Low-quality evaluation

The promotion evaluation used eleven additional deterministic low-quality views over all 3,912 calibration images.

Selected fake recalls by input condition:

Protected positive-only diagnostics also improved:

These cohorts are small and positive-only, so they do not estimate balanced accuracy or broad internet prevalence.

ONNX parity

community_forensics_low_quality_fp16.onnx is 43,778,110 bytes with SHA-256 88ca8e90e5ab33e6e13887124614e14ba96d7c8cc9ecb21505b63cdc6549ff17.

The full 11,736-prediction clean/web/hard ledger produced 27 threshold disagreements between PyTorch bfloat16 inference and FP16 ONNX: 0.2301% disagreement and 99.7699% agreement. Mean absolute raw-logit difference was 0.0855; the maximum outlier was 2.1875. The export smoke test had maximum absolute difference 0.0078 against its FP16 reference.

Robustness — failed gates

The 33-condition development red-team report is marked valid: false:

• —worst declared attack was composite_05, with balanced accuracy 0.5080 and fake recall 0.0492;
• —per-image worst-of-all-attacks fake recall was 0.0026;
• —external generative-laundering, physical-recapture, and platform-laundering cohorts were absent.

composite_05 inserts only 5% synthetic foreground into an otherwise real image. This whole-image classifier is not a localization detector. Do not interpret the low-quality promotion pass as adversarial robustness.

Intended use and limitations

Use for local, assistive screening of whole images. Do not use the score as the sole basis for accusations, moderation sanctions, employment decisions, or legal conclusions.

Known limitations include deliberate composite evasion, extreme multi-hop degradation, unseen future generators, screenshots/recaptures, incomplete real-world laundering coverage, development-set reuse for model selection, and numerical variation between inference runtimes.

License and lineage

The model is MIT-licensed. Dataset licenses remain source-specific.

• —Previous checkpoint: Thermostatic/community-forensics-frontier-detector-2026-08@16db135220b318d811b207db576d90368980b595
• —Original base: OwensLab/commfor-model-384@6076002bf0d9dd37537f965ee2f06f826c333b61
• —Upstream code: JeongsooP/Community-Forensics
• —Upstream paper: Community Forensics