lpcv-track3-pipeline

LPCV 2026 Track 3 - AIGC Image Detection Pipeline

Task: AI-generated Image Detection with Structured 8-Criteria Evaluation
Model: Qwen2.5-VL-3B-Instruct (3B params, mobile-optimized)
Method: 2-Stage LoRA SFT based on Forensic-Chat

---

Pipeline Overview

Why 2-Stage?

Based on SOTA research (Forensic-Chat, AIGI-Holmes):

• Stage 1 isolates artifact learning in the ViT ("seeing before reasoning")
• Stage 2 adds structured reasoning with JSON output matching competition format
• This prevents shortcut learning and improves generalization

---

Quick Start

1. Prepare Dataset

export HF_USERNAME=your_username
python prepare_dataset.py

2. Run Full Pipeline

bash run_pipeline.sh your_username

Or run steps manually:

# Stage 1: Train ViT for artifact detection
python train_stage1.py \
  --base_model Qwen/Qwen2.5-VL-3B-Instruct \
  --dataset $HF_USERNAME/lpcv-track3-stage1 \
  --hub_model_id $HF_USERNAME/lpcv-track3-stage1-lora

# Merge Stage 1 adapters
python merge_stage1.py \
  --base_model Qwen/Qwen2.5-VL-3B-Instruct \
  --adapter ./stage1_output/final \
  --output_dir ./stage1_merged

# Stage 2: Train LLM for JSON reasoning
python train_stage2.py \
  --merged_model ./stage1_merged \
  --dataset $HF_USERNAME/lpcv-track3-stage2 \
  --hub_model_id $HF_USERNAME/lpcv-track3-stage2-lora

# Final merge
python merge_final.py \
  --stage1_merged ./stage1_merged \
  --stage2_adapter ./stage2_output/final \
  --output_dir ./final_model \
  --hub_model_id $HF_USERNAME/lpcv-track3-final

3. Inference

python inference.py \
  --model ./final_model \
  --image sample.jpg \
  --output prediction.json

4. Evaluation

python evaluate.py \
  --predictions predictions.jsonl \
  --ground_truth gt.jsonl \
  --output results.json

5. Quantize for Qualcomm AI Hub

python quantize_for_qnn.py \
  --model ./final_model \
  --output_dir ./qnn_export

# Then compile via Qualcomm AI Hub:
bash ./qnn_export/compile_via_aihub.sh

---

Scoring Formula (Competition)

Where:

• DetectionScore = Accuracy of overall_likelihood (Real vs AI-Generated)
• ExplanationScore = 0.5 × CriterionScore + 0.5 × EvidenceScore
• CriterionScore = Exact match accuracy across 8 criteria (score 0/1)
• EvidenceScore = Semantic similarity of evidence text

---

Dataset

Custom data recommended: Generate additional training data by:

1. Taking real images (MS-COCO, Unsplash)
2. Generating fake images with SD 2.1, FLUX, etc.
3. Annotating with LLM (GPT-4o/Qwen-VL-72B) for 8-criteria analysis

---

Model Choice

For competition: Use Qwen2.5-VL-3B-Instruct to stay well above 15 TPS threshold.

---

Hardware Requirements

---

Key Hyperparameters

---

References

1. Forensic-Chat (arxiv:2509.25502) — 97.81% on AIGI-Holmes, 2-stage LoRA on Qwen2.5-VL
2. AIGI-Holmes (arxiv:2507.02664) — 93.2% ACC, DPO alignment for explainability
3. FakeShield (arxiv:2410.02761) — Decoupled detection + localization

---

License

Apache-2.0 (matching base model and datasets)