Diffusion model
lastmod 2026-07-19created 2023-10-1810 edits
In machine learning, diffusion models, also known as diffusion probabilistic models or score-based generative models, are a class of generative models. The goal of a diffusion model is to learn a diffusion process that generates the probability distribution of a given dataset.
A Diffusion model is a kind of generative model, an AI model that generates data that did not previously exist but is similar to the training data.
Categories
- Mercury 2 - Diffusion-based ultra-fast inference LLM
About
A Diffusion model is one of the deep generative models that create data. It uses a forward process (diffusion process) that gradually adds noise to data until the data becomes complete noise, and, in the opposite direction, a reverse process that creates data by gradually restoring it from noise. Since the Diffusion model is currently one of the generative models that work best for images, it seems very useful to know as a basic.
Forward Process
The diffusion process (Forward Process) that gradually adds noise to data
Reverse Process
The reverse diffusion process (Reverse Process) that restores the original data from noise
Diffusion-based SOTA models for industrial image generation (2025)
These are the models that serve as the starting point for industrial image generation. They are fine-tuned with LoRA/ControlNet and applied to industrial domains.
- FLUX.2 (Black Forest Labs, 2025.11) - FLUX.2 marks the leap from experimental image generation to production grade, and offers four variants: pro/flex/dev/open. The dev model is a 32B-parameter open-weight model that can run locally on consumer GPUs. It is currently the strongest open-source model.
- Stable Diffusion 3.5 (Stability AI) - Still has the broadest ecosystem (LoRA, ControlNet, IP-Adapter, etc.). It is the most practical choice for building industrial customization pipelines. There are Medium, Large, and Turbo variants that can be chosen depending on the use case.
- GLM-Image (Zhipu AI, 2025) - The first open-source industrial-grade image generation model, with a hybrid architecture that combines Autoregressive modeling and Diffusion decoding.
- Z-Image (2025) - A 6B-parameter single-stream Diffusion Transformer that achieves SOTA-level photorealistic image generation while training on real data only, and greatly reduces computational cost. It can run even on GPUs with 16GB or less, which makes it suited for edge deployment.
SOTA specialized in industrial defect image generation
- SeaS (ICCV 2025) - The latest SOTA in few-shot industrial anomaly image generation. A single unified model can generate images and masks for multiple anomaly types at the same time. It shows a large performance improvement over the earlier AnomalyDiffusion. Through a mechanism that separates (Separation) and shares (Sharing) defect attributes, it can even generate new anomaly types that were not in the training data.
- FAST (NeurIPS 2025) - Performs segmentation-oriented anomaly synthesis using Foreground-aware Diffusion with an Accelerated Sampling Trajectory. It achieves more accurate synthesis by focusing on the defect region (foreground), and is excellent at preserving the background.
- AnomalyControl (ACM MM 2025) - Generates highly aligned anomaly images using a controlled Diffusion model. Based on ControlNet, it can precisely control defect location and type, and it ensures text-image semantic consistency by learning Cross-modal semantic features.
- AdaBLDM (2024) - A model that combines the Blended Latent Diffusion Model with Online Adaptation.
- Training-Free Defect Generation (ICCV 2025) - Generates defect images without additional training using Stable Diffusion V2.1 and ControlNet.
Recommended combinations in practice
| Scenario | Recommended setup | Notes |
| Rapid prototyping | SD 3.5 + ControlNet + LoRA fine-tune | Broadest ecosystem, plenty of references |
| Highest-quality generation | FLUX.2 dev + domain LoRA | 32B parameters, open weights |
| Few-shot defect generation | SeaS (ICCV 2025) | Unified model, generates masks at the same time |
| Apply immediately without training | Training-Free Defect Gen (ICCV 2025) | No additional training needed |
| Multiscale inspection | ZoomLDM (CVPR 2025) | Supports high resolution up to 4096×4096 |
| Text-based defect control | AnomalyControl (ACM MM 2025) | ControlNet-based precise control |
| Unified generation + detection | DADNet / VLMDiff (2025) | End-to-end pipeline |
Key benchmark datasets
| Dataset | Description | Number of categories |
| MVTec AD | The standard benchmark for industrial defect detection. Two types (texture/object), 15 classes | 15 |
| VisA | Anomaly detection for industrial products with complex structures | 12 |
| MVTec LOCO | Anomaly detection including logical constraints (structural anomalies) | 5 |
| MHPSD | Surface defect dataset for medium and heavy steel plates (Blocky Scale, Striated Scale, Foreign Object, Inclusions) | 4 |
See also
- Deep learning
- GAN
- Stable Diffusion
- Machine Learning
- Computer Vision
- Industrial Inspection
- Generative Models
- Anomaly Detection
- Stable Diffusion
- ControlNet
- LoRA
- GAN
- MVTec AD
Favorite site
- Diffusion model - Wikipedia
- Diffusion model - Namuwiki
- Diffusion model explained (What is a Diffusion model? Diffusion model proof) - Yuni's Study
- 02. What is a Diffusion Model? - Milestones to Becoming an Image Generation AI Professional
- (Concept summary) Diffusion Model
- Understanding Diffusion language models through visualization - YouTube