InkjetOOD — Pretrained Models & Dataset

Thesis: Conditional Diffusion Models as Generative Classifiers for Out-of-Distribution Detection in Inkjet Print Quality Control Author: Ahmed Mohammed — MSc AI, Johannes Kepler University Linz (2026) Supervisor: Univ.-Prof. Dr. Sepp Hochreiter · Industrial Partner: PROFACTOR GmbH

---

What Is This?

This HuggingFace repository stores all artifacts for the inkjet print quality control experiments from the above thesis:

• Trained model weights (CDM + YOLO feature detector)
• Inkjet print dataset (FTI_Zer0P format: images + labels)
• All experiment results (metrics, score CSVs, figures, tables)

The code lives on GitHub: https://github.com/ahmed-3m/InkjetOOD

---

Quick Start

Option A — Evaluate the pretrained model (fastest, ~5 min)

# 1. Clone the code repo
git clone https://github.com/ahmed-3m/InkjetOOD
cd InkjetOOD
pip install -r requirements.txt

# 2. Download weights from this HF repo
python download_weights.py

# 3. Set your dataset path
export INKJET_DATA_DIR=/path/to/FTI_Zer0P_dataset

# 4. Evaluate
python evaluate.py \
    --checkpoint models/cdm_proposed.pt \
    --num_trials 100 \
    --out_dir results/eval_pretrained

Expected: AUROC ≈ 0.860 (single-split, N=266 test samples, seed=42).

---

Option B — Train from scratch (single split)

export INKJET_DATA_DIR=/path/to/FTI_Zer0P_dataset
python download_weights.py      # fetches YOLO weights

# Baseline (λ=0)
python train.py --epochs 100 --batch_size 128 --sep_loss_weight 0.0 \
    --eval_after --out_dir results/train_baseline

# Proposed (λ=0.01)
python train.py --epochs 100 --batch_size 64 --sep_loss_weight 0.01 \
    --eval_after --out_dir results/train_proposed

---

Option C — 5-Fold cross-validation (thesis main result)

export INKJET_DATA_DIR=/path/to/FTI_Zer0P_dataset

# Reproduces thesis Table 6.x: AUROC 0.8673 ± 0.023
python run_cv.py \
    --sep_loss_weight 0.0 \
    --epochs 100 --batch_size 128 \
    --num_trials 50 --n_folds 5 --seed 42 \
    --out_dir results/cv_lambda0

---

Step-by-Step: Load and Use Pretrained Weights

Step 1 — Download weights

Using the helper script (recommended):

git clone https://github.com/ahmed-3m/InkjetOOD && cd InkjetOOD
python download_weights.py

Or download manually from this repo's models/ folder:

from huggingface_hub import hf_hub_download
import torch

# Download the proposed CDM (λ=0.01, single-split AUROC 0.860)
path = hf_hub_download(repo_id="ahmed-3m/InkjetOOD", filename="models/cdm_v3_yolo_bbox.pt")
ckpt = torch.load(path, map_location="cpu", weights_only=False)
state_dict = ckpt["state_dict"]   # or ckpt["model_state_dict"]
print(f"Loaded {sum(v.numel() for v in state_dict.values()):,} parameters")

Step 2 — Load the model

from src.model import NoisePredictorV3
from src.diffusion import DiffusionSchedule

device = "cuda"
model    = NoisePredictorV3(base_channels=64).to(device)
schedule = DiffusionSchedule(schedule="cosine", device=device)

ckpt = torch.load("models/cdm_proposed.pt", map_location=device)
model.load_state_dict(ckpt["model_state_dict"])
model.eval()
print("Model loaded.")

Step 3 — Run evaluation

from src.evaluate import evaluate_cdm, save_results
# ... (set up test_loader with your data)
scores_df = evaluate_cdm(model, schedule, test_loader, num_trials=100, device=device)
save_results(scores_df, out_dir="results/my_eval")

For full data loading details, see evaluate.py in the GitHub repo.

Step 4 — Load the YOLO detector

from ultralytics import YOLO
yolo = YOLO("models/yolo_best.pt")
results = yolo("path/to/print_image.png", conf=0.3)
# Each detection is one print feature crop

---

Models

File	Description	AUROC	Params
models/cdm_v3_yolo_bbox.pt	CDM λ=0.01, base_ch=64 (proposed)	0.8603 single-split	9.33 M
models/cdm_v3_baseline.pt	CDM λ=0, base_ch=128 (thesis CV result)	0.8673 ± 0.023 CV	34.2 M
models/cdm_v3_test.pt	CDM λ=0.01, base_ch=64 (dev checkpoint)	≈0.85	9.33 M
models/yolo_best.pt	YOLOv8 feature detector (8 print features)	mAP@50=0.950	25.86 M
models/semantic_mismatch_angle_model.pt	Per-feature CDM: angle (~9.67:1 imbalance)	~0.82	8.94 M
models/semantic_mismatch_dist1_model.pt	Per-feature CDM: dist1	~0.89	8.94 M
models/semantic_mismatch_dots_model.pt	Per-feature CDM: dots (best feature)	~0.96	8.94 M

Which model should I use?

• Quick evaluation: cdm_v3_yolo_bbox.pt — the thesis "proposed CDM", λ=0.01, base_ch=64 (9.33 M params)
• Reproducing the thesis CV result: cdm_v3_baseline.pt — λ=0, base_ch=128 (34.2 M params); the 5-fold CV AUROC 0.8673 ± 0.023 comes from this wider model
• YOLO feature detection only: yolo_best.pt (25.86 M params, YOLOv8-based)
• Per-feature analysis: semantic_mismatch_*.pt — one model per feature, 8.94 M each, trained with a DiffGuard-style contrastive approach

Why does the baseline win on CV? On this small dataset (~1330 samples), the 5-fold CV shows λ=0 and λ=0.01 are statistically indistinguishable (0.8673 vs 0.8628). The single-split evaluation clearly favors λ=0.01 (+2.8 pp AUROC, −19 pp FPR@95). The thesis reports the CV result as the primary finding since it is more reliable.

---

Results

5-Fold Cross-Validation (Thesis Main Result)

λ	Mean AUROC	Std	Mean FPR@95
0.0 (baseline)	0.8673	0.023	0.563
0.01	0.8628	0.029	0.552
0.02	0.8510	0.033	0.624
0.05	0.8670	0.026	0.570

Protocol: 5-fold stratified CV, K=50 Monte Carlo trials, seed=42.

Single-Split Evaluation (N=266 test, seed=42)

λ	AUROC	FPR@95	K
0.0	0.8325	0.816	100
0.01	0.8603	0.626	100
0.02	0.8541	0.661	100
0.05	0.8553	0.529	100

Per-Feature AUROC (5-Fold CV, λ=0)

Feature	AUROC	Std
dots	0.956	0.035
dist6	0.936	0.067
dist1	0.887	0.073
angle	0.817	0.138
edge2	0.813	0.031
edge1	0.796	0.138
edge4	0.762	0.099
edge3	0.744	0.076

YOLO Feature Detector

Metric	Value
Precision	94.1%
Recall	89.1%
mAP@50	95.0%
mAP@50-95	84.7%

---

Dataset

The FTI_Zer0P dataset contains inkjet print images with YOLO-format annotations and quality labels.

Path	Contents
data/yolo_format_v2/	Full YOLO dataset (train/val/test images + labels)
data/metadata/feature_cls.csv	Per-sample metadata (feature class, quality label, template type)
data/crop_examples/	Sample 128×128 crops for 6 print features

---

Contents of This Repository

ahmed-3m/InkjetOOD (HuggingFace)
├── models/
│   ├── cdm_v3_yolo_bbox.pt       ← CDM λ=0.01 (thesis proposed)
│   ├── cdm_v3_baseline.pt        ← CDM λ=0 baseline (thesis CV result)
│   ├── cdm_v3_test.pt            ← Development checkpoint
│   ├── yolo_best.pt              ← YOLOv8 feature detector
│   ├── semantic_mismatch_angle_model.pt
│   ├── semantic_mismatch_dist1_model.pt
│   └── semantic_mismatch_dots_model.pt
│
├── data/
│   ├── yolo_format_v2/           ← YOLO dataset (images + labels)
│   ├── metadata/                 ← feature_cls.csv, yolo_dataset.yaml
│   └── crop_examples/            ← sample crops for 6 features
│
└── results/
    ├── cv_lambda_ablation/       ← 5-fold CV results for λ ∈ {0, 0.01, 0.02, 0.05}
    ├── single_split_lambda_ablation/  ← single-split ablation results
    ├── figures/                  ← all thesis figures (PNG)
    └── tables/                   ← all thesis tables (LaTeX)

---

Citation

@mastersthesis{mohammed2026inkjet,
  title   = {Conditional Diffusion Models as Generative Classifiers for
             Out-of-Distribution Detection in Inkjet Print Quality Control},
  author  = {Mohammed, Ahmed},
  school  = {Johannes Kepler University Linz},
  year    = {2026},
  type    = {Master's Thesis}
}

---

License

Code (GitHub): MIT License. Dataset (FTI_Zer0P) and model weights: Creative Commons Attribution 4.0 International (CC BY 4.0).