techjam-aigc/trace-rx-m-v2

TRACE-RX-M v2

TRACE-RX-M v2 is a binary image-level detector for purely AI-generated versus authentic images. It compares normalized DINOv2 patch tokens against a frozen authentic-only prototype memory and classifies directional reconstruction residuals plus retrieval statistics.

This repository contains the frozen TechJam 2026 shipping detector. It is not a standalone copy of the DINOv2 backbone: loading downloads the pinned public facebook/dinov2-base revision recorded in config.json.

Files

• s4_detector.pt: shipping detector heads and checkpoint metadata.
• s3_memory.pt: required frozen authentic prototype memory.
• config.json: pinned model and training configuration.
• s4_validity.json: held-out-generator validity decision.
• s3_capacity.json: authentic-memory capacity audit.
• evaluation/summary.json: clean and official-transform summary.
• evaluation/metrics_by_condition.csv: metrics for all official conditions.

Loading

Clone the public implementation and install its training dependencies:

git clone https://github.com/BenyAlbatross/techjam-aigc.git
cd techjam-aigc
uv sync --group train

Download this model repository, then reconstruct the frozen detector:

from pathlib import Path

import torch

from techjam_aigc.trace_rx_m.training import load_detector_checkpoint

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model, metadata = load_detector_checkpoint(
    Path("s4_detector.pt"),
    Path("s3_memory.pt"),
    device=device,
)

Images must use the canonical preprocessing defined by the public repository. The detector logit is positive for AIGC; sigmoid(logit) is the exported AIGC confidence score. The current checkpoint is not probability-calibrated.

Evaluation

The frozen checkpoint was evaluated on 6,091 development images under clean pixels and all 15 official TechJam transformation settings, producing 97,456 paired endpoints.

Evaluation	ROC-AUC	Average precision	Normalized pAUC@5%	Balanced accuracy
Clean	0.9035	0.9063	0.7794	0.8852
Macro across transformed conditions	0.9011	0.9043	0.7708	0.8839
Worst condition: Gaussian noise sigma 0.10	0.8936	0.8973	0.7461	0.8734

The evaluation did not use the organizer demonstration-only set or the locked split for model selection.

Important limitations

• Cross-generator generalization is not solved. Gemini Flash Image development ROC-AUC is 0.3279, despite strong performance on FLUX, SDXL, and GPT Image 2.
• The current TechJam robustness evaluation transforms neutralized 224-pixel BMP inputs rather than original source-resolution files.
• The model targets purely generated images. AI-edited and partially composited images are outside its trained scope.
• Scores are not calibrated probabilities and no production operating threshold is claimed.
• The detector must not be treated as sole evidence for moderation or provenance decisions.

Reproducibility and provenance

• Detector SHA-256: f811c4641a644e1eaed30891f9c075932a1de8680dce812adaf03c9a7daaf25e
• Authentic memory SHA-256: 71f0bf9edeedc5af21de1b22e3705560a1b0e8adef0c1312266cfff3eb59a7a2
• Backbone: facebook/dinov2-base
• Backbone revision: f9e44c814b77203eaa57a6bdbbd535f21ede1415

Training, augmentation, evaluation, and inference source code is maintained at https://github.com/BenyAlbatross/techjam-aigc.