RADIOlogy: A Single-Pass Approach to Training-Free Zero-Shot Anomaly Detection

What it claims

RADIOlogy is a training-free zero-shot anomaly detector built on C-RADIOv4, an agglomerative encoder distilled from SigLIP2, SAM3 and DINO-family teachers. The authors (Figueira, D'Amicantonio, Bondarev, Eindhoven University of Technology; arXiv, 29 October 2026) argue that encoders aligned to SigLIP's multi-head attention pooling (MAP) produce patch embeddings that align with text prompts on their own. CLIP's [CLS]-based design does not, which is why WinCLIP needs multi-scale sliding windows. On a seven-dataset industrial suite, the paper reports that RADIOlogy beats five training-free baselines (SAA+, WinCLIP, FADE, DIVAD, MRAD-TF) on image and pixel AUROC. It also reports being at least twice as fast as all of them, with only naive CLIP faster. For medical images, it swaps clinical terms into the prompts and claims state of the art among training-free zero-shot methods, without per-dataset prompt engineering.

How it works

Everything comes from one forward pass. The prompt ensemble is WinCLIP's industrial vocabulary, reused verbatim: 22 photographic templates crossed with normal and anomalous state phrases such as "flawless {}" and "{} with defect". Each prompt embedding is ℓ2-normalized before averaging. This yields one normal anchor t_n and one anomalous anchor t_a per category. The score for any unit vector v is the margin µ(v) = ⟨v, t_a⟩ − ⟨v, t_n⟩, a difference rather than an absolute similarity.

• Image score: µ applied to the global MAP embedding g. Unlike WinCLIP and FADE, it does not mix in dense features.
• Pixel map: the margin is computed per patch on SigLIP2-head embeddings. It is then refined in three steps: 1. Average-pool embeddings over a k×k window before scoring. 2. Diffuse the margin over an affinity graph built from raw C-RADIOv4 trunk features. The graph is a softmax with temperature τ and the diagonal excluded, updated as m ← αm + (1−α)Am for a few iterations, following CLIP-DINOiser. 3. Cluster SAM3-head features with spherical k-means into K parts, and score the pooled embedding of each part.
• Combination: the outputs are combined by multi-context aggregation. A training-free foreground prior then masks the background, using Otsu thresholding on the distance to the median border color.

flowchart LR
  E[C-RADIOv4 single pass] --> S[SigLIP2 head]
  E --> R[Raw trunk features]
  E --> M[SAM3 head]
  S --> G[Global MAP embedding] --> IS[Image score µ g]
  S --> P[Patch margins + k×k pooling]
  R --> A[Affinity diffusion]
  M --> K[Spherical k-means parts]
  P --> A --> AG[Multi-context aggregation]
  K --> AG
  FG[Border-color Otsu foreground] --> AG
  AG --> MAP[Anomaly map]

Evidence

The fetched text gives no numeric AUROC table. The only quantitative result is Figure 1, which plots image and pixel AUROC against latency on a log scale from 10 to 1,000 ms per image. The figure's caption states the "at least two times faster" margin, but the text does not give the exact AUROC values, the hardware or the dataset names.

How to use it

Code is at https://github.com/jacomof/RADIOlogy. The method needs a SigLIP2-compatible backbone, and C-RADIOv4 is the main one. It also needs category names for the prompt templates. Its free parameters are k, τ, α, the diffusion iteration count and the cluster count K. Their values are not in the fetched text.

Limits

The fetched text stops partway through the method section, so the results tables, ablations and stated limitations are not available here. The foreground prior assumes the background matches the image border color, which may not hold for cluttered scenes. The medical adaptation depends on a domain-wide vocabulary swap, and the actual terms are not shown in the fetched text.