DiffusionOOD — Pretrained Models & Results

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

---

What Is This?

This HuggingFace repository stores all trained checkpoints and raw results for the CIFAR-10 OOD detection experiments from the above thesis.

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

Method summary: A binary Conditional Diffusion Model (CDM) trained on one CIFAR-10 class detects out-of-distribution images by comparing reconstruction errors under two conditioning signals (ID vs OOD proxy). A separation loss pushes the two class embeddings apart during training.

OOD score = E_t[ ||ε − ε_θ(x_t, t, c=ID)||² ] − E_t[ ||ε − ε_θ(x_t, t, c=OOD)||² ]

Higher score → more likely OOD. This is Algorithm 1 from the thesis.

---

Quick Start

Option A — Evaluate with pretrained weights (~10 min)

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

# 2. Download pretrained checkpoints from this HF repo
python download_weights.py

# 3. Evaluate (CIFAR-10 auto-downloaded to ./data)
python scripts/evaluate.py \
    --checkpoint_path models/seed42_best.ckpt \
    --num_trials 10 \
    --data_dir ./data

Expected: AUROC ≈ 0.989, FPR@95 ≈ 0.047.

---

Option B — Train from scratch (single seed)

CUDA_VISIBLE_DEVICES=0 python scripts/train.py \
    --seed 42 \
    --separation_loss_weight 0.02 \
    --batch_size 64 \
    --max_epochs 200 \
    --eval_interval 10 \
    --scoring_method difference \
    --timestep_mode uniform \
    --experiment_tag thesis_seed42 \
    --wandb_mode disabled \
    --output_dir outputs/seed42

Expected: AUROC ≈ 0.987 at best validation checkpoint.

---

Option C — Reproduce the 3-seed study

bash scripts/run_three_seeds.sh

Expected: mean AUROC 0.9833 (seeds 42/123/456), individual: 0.9898 / 0.9914 / 0.9686.

---

Step-by-Step: Load a Pretrained Checkpoint

Step 1 — Download

from huggingface_hub import hf_hub_download
import torch

# Download the thesis headline model (seed=42, val AUROC 0.9873)
path = hf_hub_download(
    repo_id="ahmed-3m/DiffusionOOD",
    filename="models/main_training/seed42_best_auroc0.9873.ckpt"
)
ckpt = torch.load(path, map_location="cpu", weights_only=False)
print(list(ckpt.keys()))   # ['state_dict', 'hyper_parameters', ...]

Step 2 — Load the model

from src.model import DiffusionOODModel

device = "cuda"
model = DiffusionOODModel.load_from_checkpoint(path, map_location=device)
model.eval()
print("Model loaded.")

Step 3 — Run inference

import torch
from src.scoring import compute_ood_score

# x: (B, 3, 32, 32) float tensor in [-1, 1]
x = torch.randn(1, 3, 32, 32).to(device)
score = compute_ood_score(model, x, num_trials=10, device=device)
# score > 0 → likely OOD; score < 0 → likely ID
print(f"OOD score: {score.item():.4f}")

For full evaluation details, see scripts/evaluate.py.

---

Models

File	Description	Val AUROC	Test AUROC	Params
models/main_training/seed42_best_auroc0.9873.ckpt	3-seed study, seed=42 (thesis headline)	0.9873	0.9898	68.79 M
models/main_training/seed123_best_auroc0.9886.ckpt	3-seed study, seed=123	0.9886	0.9914	68.79 M
models/main_training/seed456_best_auroc0.9887.ckpt	3-seed study, seed=456	0.9887	0.9686	68.79 M
models/separation_loss_ablation/sep_loss_lambda_0p0_epoch79_auroc0.8025.ckpt	Baseline λ=0	0.8025	—	68.79 M
models/separation_loss_ablation/sep_loss_lambda_0p02_epoch29_auroc0.9911.ckpt	Ablation best λ=0.02	0.9911	—	68.79 M
models/separation_loss_ablation/sep_loss_lambda_0p1_epoch149_auroc0.9667.ckpt	Ablation λ=0.1	0.9667	—	68.79 M

Raw Score Tensors

File	Contents
models/raw_scores/seed42_cifar10_id_scores.pt	1000 ID scores (seed=42, test split)
models/raw_scores/seed42_cifar10_ood_scores.pt	9000 OOD scores (seed=42, test split)

Load with torch.load("seed42_cifar10_id_scores.pt") → float tensor of shape (1000,).

Which checkpoint for evaluation? Use seed42_best_auroc0.9873.ckpt — it gives the thesis headline result of 98.98% test AUROC. Use the raw scores to recompute AUROC instantly without running inference.

---

Results

Main Results (3-Seed Evaluation)

Seed	Val AUROC	Test AUROC	FPR@95
seed=42	0.9873	0.9898	0.047
seed=123	0.9886	0.9914	0.046
seed=456	0.9887	0.9686	0.122
Mean	0.9882	0.9833	0.072

Thesis headline result: 98.98% AUROC (seed=42). Verified from stored score tensors.

Separation Loss Ablation (seed=42)

λ	Best Val AUROC	Epoch	Gain vs λ=0
0.0 (baseline)	0.8025	79	—
0.001	0.9732	19	+17.1 pp
0.01	0.9869	—	+18.4 pp
0.02	0.9911	29	+18.9 pp
0.05	0.9851	19	+18.3 pp
0.1	0.9667	149	+16.4 pp

Thesis reports: +18.8 pp separation loss gain ✅

K Ablation (Monte Carlo Trials, seed=42)

K	AUROC	Time/sample
1	0.9100	0.010 s
5	0.9724	0.049 s
10	0.9819	0.097 s
25	0.9852	0.243 s
50	0.9864	0.486 s
100	0.9869	0.972 s

K=10 is the thesis default — best accuracy-efficiency trade-off.

External OOD Generalization (seed=42)

Dataset	AUROC
CIFAR-10 (within-split)	0.9898
Food101	0.9927
CIFAR-100	0.9697
STL-10	0.9521
FashionMNIST	0.9403
Textures	0.9284
SVHN	0.9050

---

Architecture

CIFAR-10 image (32×32×3) + noisy version x_t
            │
    ┌───────▼────────────────────────────────┐
    │  UNet2DModel (HuggingFace Diffusers)   │
    │  block_out_channels: (128, 256, 256, 256) │
    │  Attention: at 16×16 resolution         │
    │  Class conditioning: 2 embeddings       │
    │  (c=0: ID class,  c=1: OOD proxy)       │
    └───────┬────────────────────────────────┘
            │  predicted noise ε̂
    ┌───────▼────────────────────────────────┐
    │  Algorithm 1 Scoring (K=10 trials)     │
    │  score = mean_t[e(x,t,c=0) − e(x,t,c=1)] │
    │  e(x,t,c) = ||ε − ε̂(x_t,t,c)||²       │
    └────────────────────────────────────────┘

• Parameters: 68.79 M
• Diffusion schedule: Cosine cap (squaredcos_cap_v2), T=1000
• Training: 200 epochs, AdamW (lr=1e-4), batch=64, AMP 16-bit
• Separation loss: pushes ID/OOD class embeddings apart; λ=0.02 is optimal

---

Repository Contents

ahmed-3m/DiffusionOOD (HuggingFace)
├── models/
│   ├── main_training/
│   │   ├── seed42_best_auroc0.9873.ckpt     ← thesis headline model
│   │   ├── seed123_best_auroc0.9886.ckpt
│   │   └── seed456_best_auroc0.9887.ckpt
│   ├── separation_loss_ablation/
│   │   ├── sep_loss_lambda_0p0_epoch79_auroc0.8025.ckpt
│   │   ├── sep_loss_lambda_0p02_epoch29_auroc0.9911.ckpt
│   │   └── sep_loss_lambda_0p1_epoch149_auroc0.9667.ckpt
│   └── raw_scores/
│       ├── seed42_cifar10_id_scores.pt      ← 1000 ID scores
│       └── seed42_cifar10_ood_scores.pt     ← 9000 OOD scores

---

Citation

@mastersthesis{mohammed2026diffusionood,
  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}
}

---

Companion Repository

The InkjetOOD companion applies the same CDM approach to industrial inkjet print quality control:

• GitHub: https://github.com/ahmed-3m/InkjetOOD
• HuggingFace: https://huggingface.co/ahmed-3m/InkjetOOD

---

License

MIT License — see the GitHub repo for the LICENSE file.