💻 computer science

InvAD: Inversion-based Reconstruction-Free Anomaly Detection with Diffusion Models

The paper proposes InvAD, a novel inversion-based anomaly detection method that achieves state-of-the-art performance and a 2x inference speedup by circumventing explicit reconstruction and computationally expensive denoising, instead inferring latent variables via a few DDIM inversion steps to measure deviations from a known prior distribution.

Original authors: Shunsuke Sakai, Xiangteng He, Chunzhi Gu, Leonid Sigal, Tatsuhito Hasegawa

Published 2026-03-13

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Shunsuke Sakai, Xiangteng He, Chunzhi Gu, Leonid Sigal, Tatsuhito Hasegawa

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Problem: The "Perfect Copy" Trap

Imagine you are a security guard at a museum. Your job is to spot fake paintings among the real ones.

For a long time, the best way to do this was to use a magical machine (a Diffusion Model) that had only ever seen real paintings.

The Old Method (Reconstruction): You take a painting (even a fake one) and feed it into the machine. The machine tries to "clean up" the noise and reconstruct what it thinks the painting should look like based on its memory of real art.
The Score: If the machine's "cleaned" version looks very different from the original fake painting, you know it's a fake.

The Catch: This process is incredibly slow. It's like asking the machine to slowly repaint the entire canvas, pixel by pixel, thousands of times to get it right. It's also finicky; if you ask the machine to "clean" too much, it might accidentally fix a real scratch on a real painting, making you think it's a fake. If you don't clean enough, it might miss the fake.

The New Solution: InvAD (The "Stress Test" Approach)

The authors of this paper, InvAD, say: "Why are we trying to rebuild the painting? Let's just see how the painting reacts to stress."

They propose a completely different strategy: Detection via Noising.

Instead of trying to clean the image (denoising), they do the opposite. They take the image and add noise to it using the machine's logic, but they do it in a special "hidden language" (latent space) rather than looking at the actual pixels.

The Analogy: The "Mold Test"

Imagine you have a perfect, healthy apple (a normal image). You also have a rotten apple (an anomaly).

The Old Way (Reconstruction): You try to magically turn the rotten apple back into a perfect apple. If you fail to make it look perfect, you know it was rotten. But this takes a lot of time and magic energy.
The InvAD Way (Inversion): You take the apple and ask, "If I were to turn this apple into a pile of mush (pure noise), how much effort would it take?"
- The Real Apple: Because it follows the rules of nature, it turns into mush very smoothly and predictably. It fits the "mush pattern" perfectly.
- The Rotten Apple: Because it's weird and broken, when you try to turn it into mush, it resists. It doesn't fit the pattern. It ends up in a weird, chaotic spot in the "mush universe."

InvAD measures how "weird" the result is. If the result is chaotic, it's an anomaly. If it's smooth, it's normal.

Why is this a Game-Changer?

Speed (The Turbo Boost):
The old method had to take thousands of tiny steps to clean the image. InvAD only needs 3 steps to "stress test" the image.
- Analogy: The old method was like walking across a field to check the grass. InvAD is like jumping in a helicopter and looking down. It's 2x faster than the next best method, and in some cases, nearly 100x faster than older diffusion methods.
No Tuning Required:
The old method was like a radio with a sensitive dial. You had to fiddle with the "noise strength" knob perfectly to get a good result. If you turned it too far, you got static; too little, and you missed the signal.
- InvAD: It's like a digital radio that just works. Because it looks at the pattern of the noise rather than the specific pixels, it doesn't care about the exact settings. It just works out of the box.
Accuracy:
Despite being faster and simpler, it actually finds more defects than the slow, complicated methods. It catches the "rotten apples" that the old methods missed because they were too busy trying to "fix" them.

How It Works (The Secret Sauce)

The paper uses a concept called DDIM Inversion.

Think of a movie playing in reverse. Usually, diffusion models play the movie forward (adding noise) and then try to play it backward (removing noise) to create images.
InvAD takes a normal image and plays the "forward" movie (adding noise) very quickly.
Because the AI was only trained on "normal" things, it knows exactly how a normal image should turn into noise.
If the image is weird (an anomaly), the AI gets confused during this noise-adding process. The final result is "out of place." InvAD spots this out-of-place feeling and flags it as a defect.

The Bottom Line

InvAD is a new way to find defects in images (like cracks in car parts or tumors in X-rays) that is:

Faster: It skips the slow "rebuilding" process.
Smarter: It doesn't need constant adjustments.
More Accurate: It finds more problems than the competition.

It changes the question from "Can we rebuild this perfectly?" to "Does this fit the pattern of normal things?" By asking the simpler question, they solved the problem much faster and better.

1. Problem Statement

Context: Image anomaly detection (AD) is critical for industrial defect detection and medical diagnosis. Recent approaches utilize Diffusion Models to model the distribution of "normal" data.
The Limitation of Current Methods:

Paradigm: Existing diffusion-based AD methods follow a "detection via denoising in RGB space" paradigm. They add noise to an input image and then iteratively denoise it to reconstruct a "normal" version. The anomaly score is the difference (e.g., MSE) between the original and the reconstruction.
Drawbacks:
1. Noise Sensitivity: Performance heavily depends on fine-tuning the noise strength (timestep). Too much noise corrupts normal regions (false positives); too little fails to mask anomalies.
2. Computational Cost: High-quality reconstruction requires multi-step iterative denoising, leading to slow inference speeds (often < 2 FPS), which is impractical for real-time applications.
3. Trade-off: There is a fundamental tension between reconstruction fidelity and inference efficiency.

2. Methodology: InvAD

The authors propose InvAD, a novel framework based on "detection via noising in latent space." Instead of reconstructing the image, InvAD directly infers the latent representation of the input image under the learned normal distribution.

Core Concept: Latent Inversion

Reconstruction-Free: InvAD bypasses the generation of a reconstructed image ( $\hat{x}_0$ ).
Inversion Mechanism: It treats the AD task as tracing the Probability Flow Ordinary Differential Equation (PF-ODE) trajectory in reverse. Given an input image $x_0$ , the model traces the path forward to the final noise state $x_T$ (or $z_T$ in feature space).
Anomaly Scoring:
- Normal Images: Map to high-density regions of the prior noise distribution (standard Gaussian).
- Anomalous Images: Map to low-density regions because they deviate from the learned normal distribution.
- The anomaly score is derived from the deviation (typicality) of the inferred latent variable under the prior distribution.

Technical Implementation

Feature Space Diffusion:
- Instead of operating in pixel space, InvAD extracts features using a pre-trained backbone (e.g., EfficientNet-B4).
- The diffusion model is trained on these high-level semantic features ( $z$ ), making it robust to low-level variations (lighting, noise) and more efficient.
Accelerated Inversion (Few Steps):
- Standard inversion requires many steps to approximate the ODE. InvAD uses the Euler method with a very small number of steps ( $S$ ).
- The authors select a subset of timesteps (e.g., $S=3$ ) to noise the clean image rapidly.
- Key Insight: While few steps reduce reconstruction precision, they are sufficient for anomaly detection. Anomalous pixels still map to low-density regions even with coarse inversion, allowing for high accuracy without the cost of full reconstruction.
Scoring Scheme:
- To avoid the "reverse-scoring problem" (where log-likelihood fails for small anomalies), InvAD uses a hybrid score:
  - Norm-based Difference: Computes the Euclidean norm of the feature vector for each pixel. The image-level score is the difference between the maximum and minimum norms in the image. This highlights sparse anomalies while ignoring the majority of normal pixels.
  - Log-Likelihood: Combined with the norm difference to ensure robustness.

3. Key Contributions

Paradigm Shift: Identifies and addresses the limitations of the "denoising-based" paradigm, proposing a "noising-based" (inversion) paradigm that eliminates the need for explicit reconstruction.
Efficiency-Accuracy Breakthrough: Demonstrates that high detection accuracy does not require precise pixel-wise reconstruction. By using few-step inversion in feature space, InvAD achieves state-of-the-art (SOTA) performance with ~2x speedup compared to other diffusion methods (without diffusion distillation).
Plug-and-Play Design: The method is designed for the inference stage only. It can be integrated into existing diffusion-based AD models to improve their speed and accuracy without retraining the core diffusion model.
Tuning-Free: The method is robust to the selection of perturbation timesteps, removing the need for the tedious noise-strength tuning required by reconstruction-based methods.

4. Experimental Results

The authors evaluated InvAD on four benchmarks: MVTecAD, VisA, MPDD (Industrial) and BMAD (Medical).

Performance:
- Achieved SOTA Image-level AU-ROC (e.g., 99.0% on MVTecAD, outperforming OmiAD's 98.8% and DiAD's 97.2%).
- Competitive Pixel-level localization performance.
Efficiency:
- FPS: Achieved 88.1 FPS on MVTecAD, significantly faster than OmiAD (39.4 FPS) and DiAD (1.5 FPS).
- NFE (Number of Function Evaluations): Uses only 3 steps for inversion, whereas competitors often use 10–750 steps.
Generalizability:
- Successfully applied to medical imaging (BMAD), showing robustness across domains.
- When plugged into other models (DiAD, MDM), it improved their accuracy and drastically increased their inference speed.
Ablation Studies:
- Confirmed that Feature Space Diffusion and Multi-step Inversion are critical components.
- Showed that the hybrid scoring scheme (NLL + Diff) is essential for robustness against the number of inversion steps.

5. Significance and Impact

Solving the Efficiency Bottleneck: Diffusion models have been powerful but too slow for real-time industrial inspection. InvAD proves that diffusion models can be used for AD with real-time inference speeds (88 FPS) without sacrificing accuracy.
Redefining AD with Diffusion: It challenges the assumption that AD requires "reconstructing the normal." Instead, it leverages the inversion trajectory as a direct measure of anomaly.
Practical Deployment: The "plug-and-play" nature and the elimination of hyperparameter tuning (noise strength) make this approach highly attractive for deployment in automated quality control and medical screening systems.

In summary, InvAD represents a fundamental shift in diffusion-based anomaly detection, moving from a computationally expensive reconstruction task to a highly efficient, inversion-based scoring mechanism that maintains state-of-the-art accuracy.