MVTec AD · Hazelnut · Research interface

DDAD HazelnutAnomaly Laboratory

Conditioned denoising diffusion reconstructs nominal appearance, then exposes the visual evidence it cannot explain.

Open analysis workspace
LIVE SYSTEM VIEW CONNECTING
MODELDDAD · —
CATEGORYHAZELNUT
INFERENCECONNECTING

Secure inference workspace

Bring your own surface evidence.

Images are sent through a protected site endpoint to the official DDAD model. Server credentials never enter your browser.

Drop a Hazelnut image into the laboratory

or choose a PNG, JPEG, or WebP · max 10 MB

Connecting to the official DDAD inference service

Instrument readout

Analysis, without invented certainty.

SERVICE STATE
CONNECTING

Connecting to the official DDAD inference service

OFFICIAL CHECKPOINT — · FEAT —
ANOMALY SCORE
—Awaiting inference
DECISION
NOT ANALYSEDDemo threshold unavailable

Visual evidence stack

One input. Three model views. No hidden steps.

01

Input Image

EMPTY

Choose a Hazelnut image to begin

The untouched observation supplied to the reconstruction pipeline.

02

DDAD Reconstruction

PENDING

Model output will appear here

A conditioned estimate of the nominal surface generated by the official checkpoint.

03

Anomaly Heatmap

PENDING

Anomaly evidence will appear here

Pixel and feature discrepancy fused into a spatial explanation of the final score.

04

Anomaly Overlay

PENDING

Spatial overlay will appear here

The anomaly map registered directly over the observed Hazelnut surface.

How the evidence is formed

DDADseparates visual evidence from normal appearance by reconstructing what the surface should look like, then measuring what the reconstruction cannot explain.

Forward transition
q(xt∣x0)=N ⁣(αˉt x0, (1−αˉt)I)q(x_t \mid x_0) = \mathcal{N}\!\left(\sqrt{\bar{\alpha}_t}\,x_0,\,(1-\bar{\alpha}_t)I\right)
Sampled state
xt=αˉt x0+1−αˉt ϵx_t = \sqrt{\bar{\alpha}_t}\,x_0 + \sqrt{1-\bar{\alpha}_t}\,\epsilon

The input is moved to an intermediate noise state instead of being destroyed completely.

t ∼ U{1,…,T} · ε ∼ N(0,I)
Noise estimate
ϵ^=ϵθ(xt,t,w)\hat{\epsilon} = \epsilon_{\theta}(x_t,t,w)
Nominal estimate
x^0=xt−1−αˉt ϵ^αˉt\hat{x}_0 = \frac{x_t-\sqrt{1-\bar{\alpha}_t}\,\hat{\epsilon}}{\sqrt{\bar{\alpha}_t}}

A conditioned reverse trajectory reconstructs the nominal appearance while suppressing unexpected structure.

w controls the conditioned reverse trajectory
Per-pixel residual
Dp=∥y−x^0∥1D_p = \left\lVert y-\hat{x}_0 \right\rVert_1
Spatial smoothing
Dpσ=Gσ∗DpD_p^{\sigma} = G_{\sigma} \ast D_p

Local intensity differences expose fine surface changes that do not survive nominal reconstruction.

y is the observation · x̂₀ is its nominal reconstruction
Multi-scale residual
Df=∑lγl∥ϕl(y)−ϕl(x^0)∥1D_f = \sum_l \gamma_l \left\lVert \phi_l(y)-\phi_l(\hat{x}_0) \right\rVert_1
Weighted fusion
A=Df+v α DpA = D_f + v\,\alpha\,D_p

Multi-scale adapted features detect semantic inconsistencies and form the final smoothed anomaly map.

φₗ denotes adapted features at scale l · γₗ, v, α are fusion weights

Curated model evidence

Inspect genuine model evidence.

Upload a Hazelnut sample to populate this chapter with a reconstruction, heatmap, overlay, and measured decision.

GOOD CASE

Awaiting a live analysis

This slot will show genuine measured evidence after the first run.

Score
—
Threshold
—
Decision
—

System blueprint

DDAD Architecture

The complete research path from nominal diffusion training to reconstruction-based anomaly evidence.

DDAD architecture showing denoising U-Net training, feature extractor adaptation, and anomaly inference
Denoising U-Net Training

Nominal Hazelnut samples supervise time-conditioned noise prediction across the forward diffusion trajectory.

Official DDAD inference

Inspect a Hazelnut sample.

The interface will reconnect automatically when the inference service is available.