AI-Generated Image Detection on Edge

ECV Workshop @ CVPR 2026, Denver · Sponsored by Qualcomm

View code on GitHub 🤗 Model weights

The challenge

Track 3 goes beyond a yes/no classifier: the model has to decide and justify.

Saying “fake” isn’t enough — the model has to point to what gives it away.

Every prediction therefore has two parts:

• Detection — is the image Real or AI-Generated?
• Explanation — a score and written evidence for each of 8 forensic criteria:

The organizers grade the submitted binary in two stages: first the model reads the image and writes a free-form analysis across the 8 criteria (Stage 1), then it folds that analysis into one structured JSON — per-criterion score, evidence, and final verdict (Stage 2).

How it’s scored

Two constraints drive every design decision:

The accuracy score combines three measurements:

Approach

The system comes together in three parts: building a labeled dataset, a staged training curriculum, and on-device quantization & deployment.

1 · Data & annotation

Almost none of the source images came with the 8-criteria labels the task needs — so we generated them ourselves, using Qwen2.5-VL-7B as a teacher model. 89,263 images in all, balanced ~50 : 50 real / fake.

Auto-labeling — 4 questions per image

Instead of one yes/no call, we ask the teacher three vision passes, each scanning a slice of the 8 criteria, then fuse them into JSON in a final text-only pass.

Fake images get a one-line hint prepended so the model looks for artifacts; real images come out all-zero with "looks authentic" evidence — yielding a fully supervised set.

Each image becomes one annotation JSON (8 criteria shown trimmed to 3):

{
  "per_criterion": [
    { "criterion": "Lighting & Shadows Consistency",
      "evidence": "Lighting is consistent throughout; no abrupt brightness or shadow changes.",
      "aigc score": 0 },
    { "criterion": "Physical & Common Sense Logic",
      "evidence": "The fish has an unusual shape and size; the person's features look exaggerated.",
      "aigc score": 1 },
    { "criterion": "Human & Biological Structure Integrity",
      "evidence": "Unrealistic proportions and facial features; fish anatomy is inconsistent.",
      "aigc score": 1 }
    // … 8 criteria total
  ],
  "overall_likelihood": "AI-Generated"
}

2 · A staged LoRA training chain

A general VLM doesn’t know what AI artifacts look like, so we adapt Qwen2-VL-2B over four LoRA phases. Each phase trains an adapter, then merges it into the base before the next starts — so skills stack into one deployable network (LoRA+ won out over LoRA / DoRA / PiSSA; the 7B model overfit, so 2B won).

3 · Quantization & deployment

• Quantize the merged model with AIMET (W4A16) — both vision encoder and language model.
• Export through ONNX → QNN binary for the Snapdragon NPU.
• Match calibration data to the real inference distribution; a mismatch quietly wrecks quantized accuracy.

Results

Team

Team SSUPER_POWER — VIP Lab, Soongsil University:

• Dayoung Kil
• Doeon Kim
• Junyoon Lee

Tech stack

Qwen2.5-VL-7B Qwen2-VL-2B LoRA+ / DoRA / PiSSA PyTorch 2.10 LLaMA-Factory 0.9.1 AIMET Pro 1.34 (W4A16) QAIRT 2.31 · QNN

Datasets: GenImage (ADM · BigGAN) · SID-Set · SynthScars · ImageNet · COCO · ARForensics.
Code released under the MIT License.